Synergetic frequency response from multiple flexible loads

Electric Power Systems Research 145 (2017) 185–196

Contents lists available at ScienceDirect

Electric Power Systems Research journal homepage: www.elsevier.com/locate/epsr

Synergetic frequency response from multiple flexible loads Hassan W. Qazi ∗ , Damian Flynn School of Electrical and Electronics Engineering, University College Dublin, Ireland

a r t i c l e

i n f o

Article history: Received 16 March 2016 Received in revised form 15 August 2016 Accepted 3 January 2017 Keywords: Demand response Frequency control Primary reserves Thermostatically controlled loads Charging loads Synergetic control

a b s t r a c t Flexible load is likely to be a key component of future systems with the potential to enhance overall system efficiency. Multiple flexible end uses can be utilised to provide system frequency sensitive reserves. However, consideration of inter/intra day end use variability is required when tuning the flexible load response to achieve acceptable frequency behaviour. To that end, a generic synergetic mechanism is presented for frequency based primary reserve provision using representative thermostatic and charging loads. Based on differing dynamic response and load availability patterns of flexible end uses, it aims to enhance the frequency response while avoiding frequency overshoots, and minimising communication requirements. Scheme validation involves a year-long contingency analysis with varying generation mix, system demand and flexible load scenarios. The results show a marked improvement in frequency nadirs, while avoiding frequency overshoots and avoiding contracted load shedding. © 2017 Elsevier B.V. All rights reserved.

1. Introduction Recent trends in power system evolution towards a smarter grid have enabled a greater role from the demand side. Mitigating demand-generation imbalances by altering consumption patterns has received significant attention from industry and academia alike, with a number of jurisdictions utilising demand response (DR) for system ancillary service provision [1]. The imbalance following the loss of a large generator or transmission asset is traditionally resolved by conventional generation, using (frequency-based) governor droop characteristics. The same concept can be extended to flexible loads, although the latter can often respond faster to changes in frequency [2], as compared to the former, leading potentially to improved frequency nadirs. In addition, utilising flexible load instead of (or combined with) conventional units to provide primary frequency reserve (PFR) can reduce part-loading of generation, leading to improved operational cost efficiency [3]. Discretionary loads are ideal candidates for PFR provision due to negligible implications for short-term energy deferral. Such loads are abundant in the residential sector, which makes up a significant (≈30% in EU-27 countries) portion of total electricity consumption. Various individual end uses, such as fridge/freezers [4], domestic water heating [5] and plug-in electric vehicles [6], among others, have been studied in the literature for PFR provision. In the future, however, the responsive demand portfolio will likely con-

∗ Corresponding author. E-mail addresses: [email protected] (H.W. Qazi), damian.fl[email protected] (D. Flynn). http://dx.doi.org/10.1016/j.epsr.2017.01.007 0378-7796/© 2017 Elsevier B.V. All rights reserved.

sist of multiple flexible loads, as proposed by ENTSO-E (European Network of Transmission System Operators for Electricity) [7]. A portfolio of flexible end uses (as opposed to one end use) with varying seasonal, daily and weekly demand deployed in concert can result in a more uniformly available source of DR reserves. It, however, also presents challenges regarding management of the combined demand response, resulting from an autonomous “fit and forget” philosophy [8]. PFR provision from multiple end uses has recently received some attention in literature. For example, the authors in Moline-Garcia et al. [9] suggest triggering multiple flexible end uses sequentially, based on time-frequency characteristic. However, individual end use variability and its impact on control settings is not considered. A multi-step adaptive frequency restoration process, using step-wise activation of responsive demand, is highlighted in Chang-Chien et al. [10]. The required DR volume is based on real-time estimation of event severity using rate of change of frequency measurements, but the study does not consider the varying dynamic frequency response characteristics of different load types. In Vedady Moghadam et al. [11], the consequences of triggering large (varying) flexible load volumes with fixed frequency controls settings, in terms of frequency overshoots are highlighted. Staggering individual load responses into discrete time intervals is proposed as a solution, but it compromises the frequency nadir improvement. Managing the variability of flexible load resource volume by its pre-restriction in advance of an event, has also been proposed. In Weckx et al. [12] a market-based approach, with a price-based droop characteristic for flexible loads is proposed, requiring regular adjustment and updated broadcast, with multiple entity (customer, aggregator and system operator), two way communication. Study in Zhao et al. [13] proposes eval-

186

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

uation of contingency volume at load level, using a simplified power system model and regular local area communication to determine the aggregate PFR requirement. This approach however, requires sophisticated load controllers, a significant communication overhead and high computation time (up to 3.5 s) before a response, resulting in compromised nadir improvement. A completely centralised approach is presented in Biegel et al. [14] where a portfolio of flexible loads transmits state information and receiving the activation commands from the aggregator. Apart from the longer response times resulting from such a control philosophy, the flexible loads are represented as energy storage with a constant drain rate, ignoring the stochastic load behaviour based on user impacts and weather conditions, etc. All the aforementioned approaches [9–14] treat the available portfolio of flexible loads as binary (on/off) devices. The capability of charging loads, such as electric vehicles and storage space heaters, to dynamically modulate their charging rates [15] has not been utilised. Additionally, these studies are idealised, validated only on a single operating condition using reduced power system models and without the recognition of DR based PFR provision on generation dispatch. This work presents a generic synergetic control mechanism for a diverse demand resource consisting of various load groups, for primary frequency reserve provision. Detailed physical based models of charging loads (CL) and thermostatically controlled loads (TCL) are presented to highlight diurnal, weekly and seasonal variability of a flexible portfolio (Section 3). Differences in their dynamic response under autonomous control are analysed (Section 4). The variability of individual loads and their respective dynamic response characteristics, are considered while designing the synergetic control mechanism for improving frequency nadirs, while minimising communication overheads and addressing frequency overshoot problems (Section 5). Instead of a completely centralised or decentralised control philosophy, a hybrid approach is adopted, with centralised commands issued prior to an event to configure flexible load for a real-time autonomous response. The effect of static reserve (pumped hydro & interconnectors) and individual generator dynamics are considered using a detailed power system model, enabling multiple operating conditions and flexible load magnitudes to be considered. Dynamic contingency analysis is performed for a wide range of scenarios spanning a year long duration (Section 6).

2. Synergetic load response problem Flexible residential load consists of various end uses, which can be categorised as charging and thermostatically controlled loads. For the latter, electricity consumption and appliance function (i.e. temperature regulation) have a tight temporal coupling. In contrast, energy storage (charging) and later utilisation (subject to user requirements) are decoupled for charging loads. Multiple load groups (LGs), such as fridge/freezer load (FRL) and domestic water heating (DWH) represent TCLs, while CLs include electric vehicles (EVs) and storage space heating (SSH). Individual load groups exhibit varying levels of daily, weekly and seasonal variability. A potential shortage of DR-based reserve from a single group owing to its variability can be mitigated by combining other load groups, providing PFR in concert. Moreover, the urgent nature of PFR provision implies that a real time external (centralised) control is undesirable. PFR provision from various flexible loads in a real-time autonomous manner can ensure sufficient volume and speed of reserve provision. The aggregated DR load will show diurnal, weekly and seasonal variation, however a complete autonomy of response without pre-real time configuration implies minimal control over reserve variation. Hence passive settings for decentralised reserve responsiveness can result in a sub-optimal nadir improve-

ment when DR volume is relatively small. Conversely, aggressive settings, coupled with a large DR volume can cause frequency overshoots. A synergetic load control strategy, catering for varying system conditions, flexible load group variability patterns and their frequency response characteristics is therefore required. To that end, periodic centralised control may be required, but centralised communication must be minimal and inactive during a contingency. To achieve such an over-arching framework, quantitative assessment of individual end use variability is required. Also, the frequency response characteristics of CLs and TCLs need to be evaluated and exploited. Mechanism evaluation for multiple DR levels and system conditions is required to ensure generality. 3. Power system and flexible load modelling Physical based modelling of multiple load groups is required to establish diurnal, weekly and seasonal variability patterns and quantify the aggregate DR resource. The developed load group models are then integrated with a detailed system model (Irish power system in 2020) to inspect the dynamic response characteristics of individual load groups, and subsequently to develop and validate the synergetic control scheme. 3.1. Power system model The future (2020) Irish system is a relatively small system with limited DC connection (1000 MW) to Great Britain through two interconnectors, and consists of combined cycle gas turbines (4292 MW capacity), coal-fired plant (1323 MW), open cycle gas turbines (1192 MW), pumped storage hydro (292 MW), combined heat and power (161 MW), and wind farms (5 GW installed). The system model is based on a feedback loop, whereby the system frequency is calculated from the power imbalance between demand and generation, and stored energy of the rotating masses in the system [16]. All generators are assumed grid code compliant with a 4% droop setting, and individual plant characteristics, such as plant inertia, are based on data provided by the manufacturers. Wind production is assumed invariant during the POR (primary operating reserve) time frame. Frequency traces from various contingencies provided by the system operator have been used to validate the model [16]. Flexible and inflexible loads are both included. The former incorporate inherent frequency sensitivity, but do not alter their operating cycles during a disturbance. Physical-based models are adopted to better represent stochastic user behaviour and underlying load dynamics, leading to a more realistic analysis of DR actions [17]. Individual appliances for each load type are modelled, as detailed below, before being aggregated to system level using a bottom up approach. 3.2. Fridge/freezer load (FRL) Domestic cold load is modelled as the energy balance within individual appliances for better insight into their load states. The appliance model is adopted from Short et al. [4], with the addition of appliance diversity and stochastic user behaviour [18]. Different appliance components, such as freezer box, fridge air space, fridge and freezer contents are interconnected, exchanging heat with each other. A hysteresis-based thermostat maintains the cavity temperature within a defined range. For appliance i, the temperature Tn,i of the nth component is calculated as:

Nc   dTn,i = Unc,i Anc,i Tn,i − Tc,i /Sn,i mn,i dt c=1

(1)

where Nc is the number of appliance components adjacent to component n, and Unc,i and Anc,i are the thermal conductivity and heat

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196 Table 1 Fleet characteristics. Parameter

Range

Fridge freezer load

Power rating (W) Gross volume (m3 )

120–400 0.13–0.4

Electric vehicle load

Charge rating (kW) Battery size (kWh)

3.7–22 16–60

Domestic water heating load

Element rating (kW) Tank volume (l)

2.5–6 150–400

Storage space heating load

Element rating (kW) Storage capacity (kWh)

1.56–2.76 10.9–19.3

link area between adjacent components n and c. Aggregate systemwide cold load model is based on a fleet of 10 reference devices with varying ranges for power rating, gross capacity, and coefficient of performance. The mass of the freezer box and fridge airspace are set proportional to the appliance gross volume, while the specific heat capacity and food mass are varied probabilistically from values in Short et al. [4]. The power rating and gross volume range for the appliance fleet are shown in Table 1. The frequency of fridge door openings varies with time of day and is based on field trial [19], while the heat energy Eop,i corresponding to each fridge opening is represented as Eop,i = Sair air Tamb,i Vcav,i i

(2)

Sair and air are the specific heat capacity and air density, Tamb,i the ambient temperature for an individual appliance, and i the fraction of appliance cavity volume Vcav,i replaced by ambient air for each opening. Fig. 1(a) shows the fridge/freezer load profile on a summer weekday, summer weekend, winter weekday and winter weekend. The profiles exhibit a limited seasonal and weekly variability, suggesting a relatively uniform availability of flexible fridge/freezer load across the year. 3.3. Domestic water heating (DWH) load Individual water heaters, with corresponding water draw patterns, are simulated and their power consumption combined to create the overall profile. Instead of a regression approach, each water heater is represented by a classical single node model [8], based on an energy balance within the water tank. The water temperature Ttw,i in tank i is governed as:

  dTtw,i = Pi − Ui Ai Ttw,i − Tamb,i − Qevent,i /mtw,i Ctw dt

(3)

with Pi being the heating coil rating, Ui and Ai are the tank conductivity and area, mtw,i is the water tank mass, Ctw the specific heat capacity of water, and Qevent,i the energy balance change due to a water draw event. DWH power consumption depends on hot water usage, e.g. hand/dish wash, shower and bath. Events are modelled by relevant flow rates, probabilities of occurrence and seasonal variations based on metered data [20]. The influence of a water draw event on energy balance in the tank follows as



Qevent,i = wat V˙ k,i Ctw Ttw,i − Tinlet,i



(4)

where wat is the water density, V˙ k,i the volumetric flow rate corresponding to the kth draw event for the ith appliance, and Tinlet,i is the temperature of cold water entering the tank. Appliance heterogeneity and the relationship between tank size and number of occupants were established by splitting the domestic sector into dwelling types, with different occupancy ranges. Actual occupancy is evaluated probabilistically based on (morning) travel data in probabilistically based on (morning) travel data in Brown [21]. Characteristics such as tank volume and element rating depend on

187

the dwelling type, while thermostat settings, appliance efficiency and thermal capacitance are probabilistically distributed across the DWH population. The model recognises intra-day, seasonal and weekday/weekend variations in load as shown in Fig. 1(b), where there is a significant intra-day and weekly variability in DWH profiles. DWH load availability is typically limited during the night, implying a negligible response from DWH in the event of a loss of infeed event. 3.4. Electric vehicle (EV) load EV load modelling has been subdivided into a transport model and a charging decision making algorithm. The diversity in the appliance fleet, considering all-electric range, battery size and maximum power draw capabilities are incorporated. Access to various charger types, depending on plug-in location (work, leisure or home), is also modelled. In addition, vehicles are divided into commuter (CVs) and non-commuter vehicles (NCVs), with the former taking regular trips at a similar time every weekday. User trips are distributed into commuter (CT) and non-commuter trips (NCT). The generic EV load model considers Ireland as a test case. Based on Caulfield [22], 60% of EVs are assumed to be commuter vehicles. 3.4.1. Transport model EV fleet diversity is maintained by categorising 7 reference vehicle types constituting various battery and plug-in hybrid vehicles. It is assumed that 20% of the fleet can undertake 3-phase AC charging. All vehicles charge at home with a single- phase 3.7 kW charger, while work and leisure chargers can either be single or 3-phase, with ratings of 7 and 22 kW. The total number of vehicle trips, trip speed, duration and trip distance, and commuter trip times are distributed across the EV fleet by sampling a Weibull distribution based on travel survey data [21]. It has been shown previously that transport model parameters, e.g. trip distance, can be modelled, with little impact on accuracy, by Weibull, gamma, and lognormal distributions [23]. Journey departure times are distributed for each day, giving due consideration to the nature of the trip (CT or NCT) and vehicle type (CV and NCV). CT times are distributed based on available travel survey data [21], with each departure and return time slot allocated to a fraction of vehicles. The NCT journey times are generated by a normal distribution within dynamic bounds, calculated after each trip by considering time spent on previous trips. In order to reflect user behaviour, whereby departure time and stop duration for a NCT depend on previous trips, and to avoid trip overlapping, NCT departure and arrival times are dynamically updated. 3.4.2. Decision making model Instead of all vehicles charging only after the last trip of the day, charging at work and leisure trips is also considered. A decision algorithm represents stochastic charging decisions, where, as a first step, a charging decision is made, before the charge rate (fast/normal) is decided. The battery SOC upon arrival at a destination is evaluated, with charging always initiated if a charger is available and the SOC is less than a low battery threshold Bi,Lo (considered to be 30% for leisure and 50% for work charging). If the SOC is above Bi.Lo a charging decision is taken, with the charging probability higher for longer anticipated stop times and lower initial SOC:

Ui,c =

⎧ ⎪ ⎪ ⎨ (1 − SOCi ) ⎪ ⎪ ⎩

1 1 − Bi,Lo

(1 − SOCi )





1 1 − BLo

Tst,i Tstmax





ifTst,i ≤ Tstmax

ifTst,i > Tstmax

(5)

188

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

Fig. 1. Seasonal, weekly & diurnal load variability for (a) fridge/freezer load (b) domestic water heating load.

where Ui,c is the charge probability for the ith vehicle, Tst,i the expected stop time and Tstmax the cut-off maximum stop time for charge probability evaluation. Once a charging decision Dc,i is taken, a further decision, i.e. fast charging or otherwise, is required, provided fast charging capability is available. In order to minimise battery detrimental effects, it is assumed that a fast charge is less likely for shorter stop times and higher battery SOC. Fast charging probability Uri,f is given as

Ui,f =



⎧ ⎨ 1 − SOCi Tst,i × Dc,i ifTst,i ≤ T max ⎩

st

Tstmax

(1 − SOCi ) × Dc,i ifTst,i >

(6)

Tstmax

The effects of space conditioning (if activated) on EV range are considered by including an additional 20% depreciation in range [24]. It is also assumed that vehicles are pre-conditioned before a trip during hot/cold weather, which, in line with Barnitt et al. [24], is simulated by a 15 min pre-trip charge at rated power. EV owners will likely utilise dual-tariff schemes by shifting. 3.5. Storage space heating (SSH) load Storage space heaters make use of base load electricity, usually operating on dual tariffs. The energy is stored in the form of heat to be utilised during one or more bursts of heating at a later time, and are therefore analogous to a thermal battery. Modern SSH can also modulate their charging rates [15]. Space heating demand is influenced by building envelope characteristics, outside air temperature, occupancy patterns, natural ventilation and other internal heat gains. The space heating demand is based on the model developed in Neu et al. [25], with different dwelling archetypes simulated in EnergyPlus using Markov chain Monte Carlo techniques applied to Irish national time of use survey activity data for developing activity profiles. The bottom-up approach yields average energy demand for each dwelling archetype across a yearlong scenario. Due to their smaller size, compared to boiler-based central heating systems, it was assumed that storage heaters are used in flat-type dwellings only. Storage heater sizes were allocated based on the energy requirement per dwelling. The charging profile, assuming a dual tariff, results in an aggregate load shape that normally peaks

as the off-peak tariff period begins and progressively reduces as appliances satisfy their respective charging needs. SSH load has, therefore, been modulated to more evenly distribute the resource availability, for later DR utilisation across the off-peak window Th , by adjusting the power draw rate, PiSSH , for the ith appliance as follows, where Ei,h is the heat energy required per hour for the next day, Ei,0 residual energy and Hl,i is the hourly loss rate. PiSSH =

24  h=1



Ei,h 1 + Hl,i



+ Ei,0 /Th

(7)

Fig. 2 represents the daily, weekly and seasonal variations in charging load availability. The increased availability of charging load during night as opposed to water heating and fridge/freezer load is evident, suggesting the complementary nature of various flexible load group availability profiles. 3.6. Aggregate flexible load Fig. 3 shows the diurnal variability of various flexible load resources on a weekday and a weekend, during winter and summer seasons. The aggregate DR resource at night varies significantly between the seasons, primarily due to the strong SSH load during winter. Similarly, the weekday and weekend aggregate profiles are different, influenced by the delayed morning peak for the DWH profile on the weekend. The reduced fridge and water heating load availability at night is mitigated by the presence of storage heating and electric vehicle load. Although various end uses exhibit intra-day and intra-week resource variability, their combination results in an increased demand resource with sufficient availability for PFR provision across the day. A combination of flexible loads is required to maintain significant PFR levels across the year, thereby necessitating a synergetic demand control strategy. 4. Load dynamic response characteristics The dynamic response of various flexible load types (TCLs and CLs) providing reserve significantly affects the system frequency profile following a contingency. An assessment of the response characteristics of each load type providing PFR is therefore necessary. The conclusions drawn here, regarding suitable control

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

189

Fig. 2. Charging load variability (a) electric vehicle load, (b) storage heating load.

Fig. 3. Seasonal, weekly & diurnal load variability for combined flexible load.

strategies and settings for each load type, will subsequently inform the synergetic mechanism for various load groups belonging to either (TCL/CL) load type. All appliances perform local frequency measurement and respond according to pre-configured settings, as envisaged by ENTSO-E [7], thereby avoiding communication latency. A response time of 100 ms for all appliances is considered.

4.1. Thermostatically controlled load characteristics TCLs are generally binary appliances (On/Off), with their power consumption restricted to rated values when turned on. During a contingency, all responsive TCLs can be switched off through a uniform threshold value, resulting in a relatively quick arrest of a declining system frequency. However, such an approach can result in system frequency oscillations due to load over-responsiveness [11]. Solutions include mimicking the governor droop of a conventional plant, by either distributing frequency thresholds across individual appliances or altering the thermostat set points in proportion to the system frequency deviation. The latter approach is

preferable due to inherently considering user comfort, since the likelihood of appliances responding is inversely related to their respective energy requirement, i.e. warmer appliances for heating end uses (such as DWH) switch off first and recover last. This strategy can be implemented through a deadband, Db and operating range, Or with 0% of available DR triggered at the start of the operating range and 100% at its end. For individual TCLs, a proportional change between appliance thermostat setpoints and the frequency, governed by Or :

inew =

⎧ f − Db max ⎪ ×  + iold ifDb ≤ |f | ≤ Db + Or ⎪ i ⎪ Or ⎨ ⎪ ⎪ ⎪ ⎩

imax + iold ifDb + Or ≤ |f |

(8)

iold Otherwise

where inew is the altered thermostat setpoint, imax is the maximum allowable setpoint deviation to maintain user comfort, and iold is the original setpoint. However, from Fig. 1(b), TCLs, such as DWH, have significant seasonal and diurnal variation. Non-

190

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

Fig. 4. (a) System frequency, (b) demand side reserves with uncontrolled TCLs and CLs, and proposed TCL control.

dynamic (fixed Db , Or ) control, even under the control scheme represented by Eq. (8), can result in a frequency overshoot due to a high volume of DR reserve. Fig. 4 shows one such scenario where the loss of a large infeed (430 MW) with a system demand of 5654 MW, 1000 MW of wind generation and 1100 MW of available DWH reserve is considered. With a fixed Or and Db of 0.1 Hz in this case a frequency overshoot occurs, since the triggered reserve volume exceeds the generation loss. Each appliance reacts to a thermostat setpoint change, reduced in proportion to the frequency fall, interrupting the heating cycle. Subsequently, the energy consumption from an appliance is deferred until the next heating cycle, and so is not immediately affected by a restoration of thermostat setpoints, as the frequency recovers. TCL over responsiveness can be reduced by setting conservative values for Db and Or , as seen by increasing Db from 0.1 to 0.3 Hz in Fig. 4. However, in order to avoid a trade-off between reserve over-responsiveness and a lower frequency nadir, an alternative mechanism, based on pre-configuration of the responsive reserve volume is proposed: a global control signal ˛ (t) is calculated which represents the target fraction of available reserve to be activated

˛ (t) =

⎧ G ⎨ Rreq (t) , ∀Pag (t) > 0 ⎩

G Pag (t)

(9)

G 1, ∀Rreq (t) > Pag (t)

G The available reserve for a certain load group ‘g’, Pag (t), can be forecasted using detailed stochastic flexible load models, such as those presented in Section 3. The scalar value ˛ (t), between 0 and 1, is globally broadcasted and periodically altered based on a change in unit commment schedule. The global control signal, supported by pre-set Db and Or values, is implemented probabilistically on board each appliance to determine its later participation in PFR provision. The proposed control scheme is applied to the same system configuration as before, labelled as “controlled TCL” in Fig. 2, where even a narrow Db of 0.1 Hz does not compromise the frequency nadir. A frequency overshoot is avoided due to only a fraction of available DR

being activated. It is noteworthy that since the mechanism depends on the PFR requirement, based on the largest infeed, it can still result in a frequency overshoot for smaller trips, particularly if Db and Or are tighter to the nominal frequency. The proposed decentralised scheme avoids real-time communication, since ˛ only varies with (significant) unit commitment changes. Since the control signal ˛ is the same for each load, the signal can be globally broadcasted, reducing the need for dedicated communication channels and the volume of data transmitted. 4.2. Charging load characteristics In contrast with TCLs being binary appliances (On/Off), CLs can modulate their power draw to an arbitrary value between zero and rated consumption [15,26]. CL control can adjust in a manner akin to droop, using a Db and Or , with appliance consumption varying linearly across the defined Or as:

Pinew =

⎧ f − Db pre ⎪ 1− × Pi ifDb ≤ |f | ≤ Db + Or ⎪ ⎪ Or ⎨ ⎪ ⎪ ⎪ ⎩

0ifDb + Or ≤ |f |

(10)

Pirat Otherwise

Pinew is the adjusted power draw in response to a DR event and pre Pi is the pre-DR event power draw for the CLs. In the case of a DR event, the thermostat setpoints for TCLs change in proportion to the frequency drop, interrupting the heating cycle. Energy consumption is then deferred until the start of the next heating cycle, despite later restoration of the setpoints. In contrast, since CLs are not cyclic appliances, they continuously track the system frequency and modulate their power consumption accordingly. This key difference results in CLs showing improved frequency tracking capabilities. Such self-regulation of the DR magnitude avoids the need for a global control signal dispatch, as required in the TCL case. As a further addition to the test conditions of Fig. 4, 1100 MW of CLs, labelled as “uncontrolled CL”, provide PFR with a Db and Or of

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

0.1 Hz. A frequency overshoot is avoided without compromising the frequency nadir. It must be noted that the pre-dispatch procedure Eq. (9) is not applied for CLs due to their superior frequency tracking. Additionally, the dynamic regulation capability of CLs allows tighter Db settings (upto 0.05 Hz) without risking a frequency overshoot. Tighter Db values are likely to result in improved frequency nadirs. Analysis of TCL dynamic characteristics reveals a trade-off between nadir improvement and frequency overshoots, which can be addressed through the proposed mechanism Eq. (9), involving periodic communication of a global control signal. In contrast, CLs, with their ability to self-regulate, provide improved frequency nadirs while avoiding overshoots without requiring exogenous control. CLs exhibit better dynamic characteristics; however, PFR provision from a number of load groups requires additional characteristics to be considered, in a synergetic control mechanism. 5. Load control philosophy The aim of synergetic control is to ensure rapid deployment of available flexible load, thereby ensuring nadir improvements, while avoiding over-responsiveness which may result in frequency overshoots. To reduce the implementation costs, the communication between a central control entity, such as an aggregator or TSO, must also be minimised. Achieving these objectives through synergetic control of various flexible loads requires an evaluation of load group characteristics. 5.1. Load availability A load group with a more uniform diurnal, weekly and seasonal availability is a more reliable source of PFR and should be triggered first. In the test system, domestic fridge/freezer loads (FFL) have excellent availability (Fig. 1) with the additional advantage of being small in magnitude relative to the PFR requirement. In contrast, DWH and EV loads have good seasonal availability, but their intraday availability is poor. Similarly, SSH are available only at night and so have poor availability. Each load group has varying energy recovery characteristics, but occurring beyond the PFR time scale. Generally, additional generation can be activated to alleviate the longer term impacts of load recovery. 5.2. Communication requirements

191

Table 2 Domestic load characteristics. Resource

Typical intra-day Seasonal variation load variation (MW of energy standard deviation) consumption (MWh/day)

Availability exceeding reserve requirement (% cases)

FRL (TCL) EV (CL) SSH (CL) DWH (TCL)

1.3 134 59 200

0 14.2 1.6 27.2

167 300 1597 670

if all available DR resource is activated. Under the synergetic mechanism, fridge load is considered first in case of a contingency, based on its high availability, and not requiring centralised communication (due to low aggregate volume). EV and SSH (charging) loads are bundled together to mitigate the low intra-day and seasonal availability of SSH, offering superior dynamic performance compared to TCLs. DWH load has high intra-day variability, requires pre-real-time dispatch (periodically) when its aggregated volume exceeds the largest infeed, and so is triggered last, as demonstrated in Table 2. For other systems, with different load types, a similar process can be followed, leading to a hierarchy of responding load groups. For the test system (Irish power system) static sources of reserve, such as HVDC interconnectors and pumped storage plant, trigger at defined frequency deviations beyond 0.4 Hz. Therefore, in order to fully exploit the benefits of fast response speed, the DR resource is utilised within a frequency deviation of 0.3 Hz from nominal. A deadband Db of 0.05 Hz is also included to avoid unnecessary triggering. Under the synergetic mechanism, the reserve provided from non-base TCL (DWH) is periodically re-scheduled, depending upon the available supply forecast. All non base TCLs are responsive by default. However, ˛ = 0 is broadcasted if the reserve requirement, Rreq can be met by base TCL (FRL) and charging loads (EV, SSH) alone. Otherwise, if the combined reserve from all load groups exceeds the PFR requirement, then ˛ is calculated as follows and broadcasted to DWH load:

⎧ 1, ∀Rreq (t) ≥ C (t) ⎪ ⎪ ⎪ ⎪ ⎨ ˛ (t) =

0, ∀Rreq (t) < C (t) ∧ Rreq (t) ≤

Nb

5.3. Dynamic performance The ability of a frequency controlled load to operate with tighter Db and Or settings, and under all contingency scenarios, without causing frequency overshoots. As seen in Section 4, charging loads can accept tighter Db and Or settings in comparison with TCLs. Moreover, due to the small volume of aggregate fridge load, it can also accept tighter Db and Or , without causing frequency overshoots. The larger magnitudes of intra-day variation and seasonal variation in daily energy consumption signify a non-uniform availability pattern. Table 2 also shows the fraction of cases across the year when the magnitude of available load resource exceeds the reserve requirement, which can potentially lead to a frequency overshoot,

(t) +

NCL

⎪ ⎪ ⎪  Nb ⎪ ⎩ R (t) − PEv (t) ∀Rreq (t) < C (t) ∧ Rreq (t) > PDWH (t)

Since synergetic control aims to avoid communication where possible, load groups requiring no communication should be activated first. From Section 4, decentralised activation of aggregate TCL volume exceeding the generation loss can cause frequency overshoots, thereby requiring pre-real-time communication of a control signal ˛. In contrast, FFL and CLs do not require communication owing to their small magnitude (compared to the PFR requirement) and frequency tracking capabilities respectively.

P base base=1 ag

where C (t) =

Nb base=1

base Pag (t) +

P CL CL=1 ag

base Pag (t) +

base=1

NCL CL=1

(t)

CL TCL Pag (t) + Pag (t)

NCL CL=1

CL Pag (t)

(11)

(12)

Nb and NCL are the number of base and charging load groups, while base , P CL , P TCL are the aggregate consumption forecast of each base, Pag ag ag CL and non-base TCL group. In order to further mitigate communications, ˛ is updated from one interval to the next only if deviation from previous value exceeds a defined threshold. 5.4. Control system architecture In line with ENTSO-E’s demand side vision [7], it is assumed that all flexible load appliances under consideration are equipped with frequency sensitive controllers with the capability to alter appliance operation (on/off) based on local frequency measurements. The installation of these controllers among consumers is likely to be undertaken by demand aggregators, in exchange for monetary benefits. The aggregator acts as an intermediary between the transmission system operator (TSO) and individual consumers, providing flexible load availability to the TSO and maintaining the

192

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

required demand based reserve levels as requested. The generation based reserve is allocated directly by the TSO in accordance with the UC/ED schedule, and the magnitude and speed of response from flexible loads can be controlled through the selection of deadband and operating range for load frequency controllers, as per Eqs. (8) & (10). Under the proposed synergetic control mechanism, Fig. 5, the aggregator, after receiving the reserve requirement from the TSO for each hour, evaluates whether a control signal broadcast is required, based on Eq. (11) and transmits ˛ (t) to the DWH load. While electric vehicle, storage space heating and fridge/freezer loads respond based on the f measurement and their respective droop control settings (Db and Or ), the response of each DWH appliance depends on its decision to participate (based on ˛ (t)) in addition to f and its droop setting. At an aggregate level, the magnitude of the DWH responsive load is thereby configured as the available load changes, to avoid frequency overshoots. Although it is possible that the response from each load group is driven solely by the aggregator through price signals as opposed to f measurement, such a price based demand response strategy is unsuitable for a fast-acting reliability service such as primary frequency reserve. The proposed control scheme exploits the inherent variability pattern (in the context of system reserve requirement), and the dynamic characteristics of each load group to minimise the external control requirement, resulting in minimal external communication. 6. Synergetic load control evaluation In order to test the effectiveness of the proposed control scheme two cases are established: I Conventional reserve case: primary frequency reserve is provided solely by generation portfolio and static reserve sources. II Demand response reserve case: primary frequency reserve is provided by the generation portfolio, static reserve and available demand resources. A year-long unit commitment/economic dispatch is obtained for both cases using GAMS unit commitment tool [8]. The unit commitment for the “DR reserve” case schedules the load groups outlined earlier for PFR provision, based on their forecast availability (Section 3). Both year-long production schedules serve as inputs to a dynamic test system model (inclusive of flexible load models, in “DR reserve” case), outline in Subsection 3.1. A loss of largest infeed contingency is introduced for each hour across the year, with the flexible load responding in accordance with the proposed load control scheme (Section 5) for “DR reserve” case. The economic impacts of the DR based reserve provision are quantified through analysing the production schedules, while the resulting frequency profiles from the “conventional reserve” and “DR reserve” cases are compared to evaluate the performance of proposed control scheme.

Table 3 System performance improvement with synergetic load control. Resource

Conventional reserve case

DR reserve case

Average nadir (Hz) Average steady state error (Hz) Average time to steady state (s) Contracted load shedding (% cases)

49.42 0.28 26.1 7.5

49.66 0.25 21.8 0

reduces the need for load shedding from 7.5% of cases in the conventional reserve scenario to 0% with synergetic DR. In addition to an improvement in the frequency nadirs, the provision of DR based reserve reduces the system steady state frequency error, signifying an improved post-event state compared to the “conventional reserve” case, Table 3. It can also be observed that the average time to reach steady state is also improved when DR based reserves are employed, owing to the faster response time of flexible load compared to conventional generation. An improvement in the time to reach steady state is indicative of a reduction of the duration of the transient period, thereby providing an improvement in dynamic performance. Generally, higher levels of available DR reserve result in improved nadirs. In addition, increased SSH and EV loads at night lead to large benefits due to tighter allocated Db and Or values. Given lower volumes of CLs during the day, the proposed Db and Or settings for non-base TCL, i.e. water heating, could be made tighter (closer to nominal), for further nadir improvement. Such an approach, however, would require time sensitive settings for each appliance, implying expanded processing capacity for the local load controllers. 6.2. Flexible load communication requirements As per the synergetic control mechanism, the non-base TCL (i.e. DWH) requires periodic dispatch if the combined reserve from all load groups exceeds the PFR requirement or the reserve requirement can be met by base TCL and CLs combined. Control signal dispatch is further mitigated by renewing the control signal from one time step to the next, if it changes by more than 5%. Fig. 7 shows the likelihood of the control signal ˛ being broadcast for typical summer and winter days. During the night and early morning hours (12 a.m. to 10 a.m.), the charging loads mainly provide the reserve, so the need to broadcast a control signal (restricting DWH reserve) is raised at the start, around 12 a.m. and infrequently updated until 7 a.m., particularly during winter owing to the presence of storage heating load. In contrast, during daylight hours the reduced availability of storage heating loads and also DWH loads lowers the likelihood of DR over-responsiveness. The broadcast probability increases after 5 p.m. due to evening EV plug-in charging load. The broadcast probability in summer months is usually lower (35% cases), as compared to winter (39% cases) due to reduced storage heating load, with an annual 63% reduction in broadcast probability, as compared to a centralised control mechanism.

6.1. Impact on system dynamic performance 6.3. Contribution to system reserve requirements Fig. 6 shows the resulting distribution of frequency nadirs for year-long dispatches with/without demand response. The latter case, using only conventional plants to provide reserve, shows peaks at 49.5 and 49.3 Hz, representing the activation of HVDC interconnector response and contract load shedding. However, by deploying the synergetic mechanism a significant shift to higher frequency nadirs is seen, with an improvement on average of 0.24 Hz at each hour, mainly due to the faster response provided by flexible load. Improving the nadir, apart from enhancing system stability, also lessens the use of static reserve sources and

Due to their time-varying availability, individual load groups can only provide a fraction of the system reserve across the day, although the combined contribution approaches 300% of the reserve requirement at certain times, Fig. 8(a). On average, however, the SSH and EV load contribute 16% and 44% of the PFR requirement across the year. Although the fridge load is much less variable, it can still only provide 10% of the PFR requirement, due to the small resource volume. With DWH load providing a further 42%, the aggregate DR resource contributes 100% of the PFR

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

193

Fig. 5. Overview of synergetic control mechanism.

Fig. 6. Year-long frequency nadir probability distributions.

requirement (on average) across the year. Hence, at certain times, the availability of 100% PFR from DR can displace reserve from conventional generation, and may reduce the number of online units, and the system cost. Fig. 8(b) shows the impact of excess DR availability on the average frequency overshoot for each hour across the day. It can be observed that during times of increased DWH load, a significant frequency overshoot is observed if the proposed control mechanism (Section 5) is not employed. Fig. 8(b) demonstrates the effectiveness of the co-ordinated load control to alleviate frequency overshoots owing to excess DR resource availability. It must be noted, however, that conventional generation is required to provide voltage control, inertial support and short circuit power,

despite 100% of the reserve requirement being met by DR at certain times. System dispatches recognise these requirements, and incorporate a constraint for dispatching at least 5 large generators, in line with the Irish TSO’s operational policy. A sample case from the above analysis is shown in Fig. 9. The system load is 5900 MW with 2900 MW of wind generation and available reserve capacity from FRL, EV, SSH and DWH loads is 35, 300, 210, 120 MW. Synergetic control significantly improves the frequency nadir, compared to the conventional reserve case. Although the available DR reserve volume exceeds the infeed loss, a frequency overshoot is avoided due to non-base TCL (DWH) predispatch and CL frequency tracking capabilities.

194

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

Fig. 7. Seasonal and diurnal variation of control signal broadcast probability for year long DR provision.

Fig. 8. (a) Available load resource in comparison with reserve target (b) impact of excessive reserve availability on frequency overshoot with and without proposed load control.

6.4. Economic considerations In addition to frequency stability improvement, the provision of primary frequency reserve from flexible load leads to a reduction of 36 million Euro in system operation cost across the year. Substitution of conventional plant based reserve by DR leads to a reduction in the number of online conventional plant, while maintaining adequate frequency stability. DR based reserve provision mitigates the impact of wind variability, as shown in Fig. 10(c), in terms of a reduction in the number of plant starts across the year, which is reduced from 1184 to 1050. The reduction in plant starts contributes to a reduction in plant maintenance cost and is likely to enhance plant life [27]. DR based reserve provision also results in reduced periods of wind curtailment, with an additional 41 GWh wind production across the year.

The costs associated with the implementation of the proposed control scheme include the capital costs, primarily associated with the price of individual “frequency” controllers, aggregator communication systems and load management platform (between aggregator and DWH appliances). The cost for an individual frequency controller has been estimated to be as low as 4$ [4], while an openADR compliant communication and load management platform has typical technology enablement costs ranging from $170/kW to $300/kW [28]. The economic benefits obtained from DR based primary frequency reserve activation can be distributed among the aggregators and consumers depending on the energy market and payment mechanisms, which will vary with the control jurisdiction [1,29].

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

195

Fig. 9. (a) System frequency, (b) reserve provision from individual end-uses in a sample scenario following the loss of 400 MW infeed.

Fig. 10. Comparison of (a) yearly operation cost (b) mean value of wind dispatch, (c) yearly plant switch on/off.

7. Conclusions Different load end uses can combine to increase the volume and availability of DR-based reserves, such that, in the future, frequency sensitive appliances from various load categories will constitute the flexible demand portfolio [7]. A synergetic mechanism has been presented for utilisation of different DR sources in concert, providing a balance between improving the frequency nadir and avoiding frequency overshoots. A generic framework prioritises DR resources, considering their individual diurnal and seasonal availability, dynamic characteristics and (external) communication requirements. Detailed physical-based models for four end uses, representing charging and thermostatically controlled loads have been developed to represent load variability, with the framework

validated by year long contingency analysis for a range of generation mix, system demand and flexible load scenarios. Through suitably blending the large but variable volume of certain load types, e.g. domestic water heating, with the superior controllability and low communication requirements of others, e.g. storage space heating, electric vehicles, the results show that the proposed mechanism avoids frequency overshoots, while significantly improving the frequency nadir (0.24 Hz on average). Utilisation of multiple load sources increases DR availability, enabling 100% of PFR requirement (on average) across a year. In addition, due to the fast DR response, contracted load shedding is avoided. The generic framework can clearly be extended to other systems, while overfrequency contingencies can also be incorporated.

196

H.W. Qazi, D. Flynn / Electric Power Systems Research 145 (2017) 185–196

Acknowledgement This work was conducted in the Electricity Research Centre, University College Dublin, Ireland with the financial support of Science Foundation Ireland under grant number 09/IN.1/I2608. References [1] F. Shariatzadeh, P. Mandal, A. Srivastava, Demand response for sustainable energy systems: a review, application and implementation strategy, Renew. Sustain. Energy Rev. 45 (2015) 343–350. [2] A. Saffarian, M. Sanaye-Pasand, Enhancement of power system stability using adaptive combinational load shedding methods, IEEE Trans. PWRS 26 (3) (2011) 1010–1020. [3] M. Aunedi, P.-A. Kountouriotis, J. Calderon, D. Angeli, G. Strbac, Economic and environmental benefits of dynamic demand in providing frequency regulation, IEEE Trans. Smart Grid 4 (4) (2013) 2036–2048. [4] J. Short, D.G. Infield, L. Freris, Stabilization of grid frequency through dynamic demand control, IEEE Trans. PWRS 22 (3) (2007) 1284–1293. [5] Z. Xu, J. Ostergaard, M. Togeby, Demand as frequency controlled reserve, IEEE Trans. PWRS 26 (3) (2011) 1062–1071. [6] Y. Mu, J. Wu, J. Ekanayake, N. Jenkins, H. Jia, Primary frequency response from electric vehicles in the Great Britain power system, IEEE Trans. Smart Grid 4 (2) (2013) 1142–1150. [7] ENTSO-E Network Code on Demand Connection, European Network of Transmission System Operators for Electricity, ENTSO-E, 2012. [8] H. Qazi, D. Burke, D. Flynn, Challenges in utilisation of demand side response for operating reserve provision, in: Innovative Smart Grid Technologies Conference Europe, Istanbul, 2014. [9] A. Moline-Garcia, F. Bouffard, D. Krischen, Decentralized demand side contribution to primary frequency control, IEEE Trans. PWRS 26 (1) (2011) 411–419. [10] L.-R. Chang-Chien, L.N. An, T.-W. Lin, W.-J. Lee, Incorporating demand response with spinning reserve to realize an adaptive frequency restoration plan for system contingencies, IEEE Trans. Smart Grid 3 (3) (2012) 1145–1153. [11] M. Vedady Moghadam, R. Ma, R. Zhang, Distributed frequency control in smart grids via randomized demand response, IEEE Trans. Smart Grid 5 (6) (2014) 2798–2809. [12] S. Weckx, R. D’Hulst, J. Driesen, Primary and secondary frequency support by a multi-agent demand control system, IEEE Trans. PWRS (2016) (in press). [13] C. Zhao, U. Topcu, S. Low, Optimal load control via frequency measurement and neighborhood area communication, IEEE Trans. PWRS 28 (4) (2013) 3576–3587. [14] B. Biegel, L. Hansen, P. Andersen, J. Stoustrup, Primary control by on/off demand-side devices, IEEE Trans. Smart Grid 4 (4) (2013) 2061–2071.

[15] H. Qazi, D. Flynn, Provision of flexibility at high wind penetration levels using modern storage heater load, in: PES General Meeting | Conference & Exposition, 2014 IEEE, Washington, 2014. [16] J. O’Sullivan, A. Rogers, D. Flynn, P. Smith, A. Mullane, M. O’Malley, Studying the maximum instantaneous non-synchronous generation in an island system—frequency stability challenges in Ireland, IEEE Trans. PWRS 29 (6) (2014) 2943–2951. [17] S. Shao, M. Pipattanasomporn, S. Rahman, Development of physical-based demand response-enabled residential load models, IEEE Trans. PWRS 28 (2) (2013) 607–614. [18] H.W. Qazi, D. Flynn, Analysing the impact of large-scale decentralised demand side response on frequency stability, Int. J. Electr. Power Energy Syst. 80 (2016) 1–9. [19] Jean-Paul Zimmermann, Matt Evans, Jonathan Griggs, Nicola King, Les Harding, Penelope Roberts, Chris Evans, 2012. Household Electricity Survey a Study of Domestic Electrical Product Usage, Intertek Testing and Certification Ltd., Available online: https://www.gov.uk/government/uploads/system/ uploads/attachment data/file/208097/10043 R66141HouseholdElectricitySurveyFinalReportissue4.pdf. [20] U. Jordan, K. Vajen, Realistic Domestic Hot Water Profiles in Different Time Scales, International Energy Agency, 2001. [21] Milward Brown, National Transport Authority National Household Travel Survey 2012, National Transport Authority, Dublin, 2013. [22] B. Caulfield, Identifying the characteristics of households with multiple car ownership, in: 90th Annual Meeting of the Transportation Research Board, Washington, 2011. [23] P. Plotz, N. Jakobsson, F. Sprei, S. Karlsson, On the distribution of individual daily driving distance, in: European Electric Vehicle Congress, Brussels, 2014. [24] R. Barnitt, A. Brooker, L. Ramroth, J. Rugh, K. Smith, Analysis of Off-Boad Powered Thermal Preconditioing in Electric Drive Vehicles, NREL, Shenzhen, 2010. [25] O. Neu, S. Oxizidis, D. Flynn, F. Pallonetto, D. Finn, High resolution space-time data: methodology for residential building simulation, in: 13th Conference of International Building Performance Simulation Association, Chambéry, 2013. [26] P. Almeida, J. Lopes, F. Soares, L. Seca, Electric vehicles participating in frequency control: operating islanded systems with large penetration of renewable power sources, in: PowerTech, IEEE, 2011, Trondheim, 2011. [27] N. Troy, D. Flynn, M. Milligan, M. O’Malley, Unit commitment with dynamic cycling costs, IEEE Trans. Power Syst. 27 (4) (2012) 2196–2205. [28] G. Ghatikar, D. Riess, M. Piette, Analysis of Open Automated Demand Response Deployments in California and Guidlines to Transition to Industry Standards, LBNL, 2014. [29] M. Behrangrad, H. Sugihara, T. Funaki, Integration of demand response resource payback effect in social cost minimization based market scheduling, in: 17th Power System Computation Conference, Stockholm, 2011.