After nearly a decade of advanced metering deployments, the power industry is ripe for major advancements in the role of demand-side management. As smart grid technologies ranging from sensors to data analytics mature and gain traction, new market models are emerging to unleash the value of information technology in electricity markets.
Before many companies began developing new products and services to capitalize on this opportunity, national laboratories within the Department of Energy (DOE) were already engaged in ambitious and promising demonstration projects to prove the value of demand participation with evidence from field data. The most prominent projects have come from the Pacific Northwest National Laboratory (PNNL)—a DOE laboratory operated by the Battelle Memorial Institute and affiliated with the GridWise Architecture Council (GWAC).
A giant first step: Olympic Peninsula Project
The PNNL’s Olympic Peninsula Project (2007) was the starting point for the development of price-driven demand participation. The field experiment, which was conducted from early 2006 through March 2007 in the Pacific Northwest, demonstrated for the first time that two-way communication between grid operations and distributed resources could enable the dispatch of automated demand-side resources based on real-time price signals. The test region, which is served by a long and essentially radial transmission system, was a prime candidate for “non-wires” enhancement solutions to its electricity infrastructure. The experiment involved 112 residential participants from two utility service territories, three municipal water pumps, and three on-site generators. Additionally, a commercial building owned by the PNNL ran a building automation system (BAS) as part of the demonstration project.
In order to make participants’ demand responsive to price signals, each home was equipped with the hardware necessary to automatically monitor and control the HVAC systems and water heaters. Dryers equipped with price signal displays were also employed, although their operation was done manually by participants. An online energy management platform was provided to participants, streamlining access to comfort settings and detailed usage information. The intelligent operation of devices was executed through a virtual interface built by IBM which, without modifying the appliances themselves, essentially created new, intelligent virtual devices that could participate in the electricity market.
Recruiting of participants
Participants were recruited through media such as a town-hall meeting, a radio show interview, newspaper advertisements, and word of mouth. Those interested were then screened based on some technical pre-qualifications. Applicants were offered a cash incentive averaging $150 based on their relative responses to energy price signals—there were no penalties for negative balances. Based on their stated preferences, participants were evenly allocated into fixed, time-of-use, and real-time pricing arrangements. At the end of the project, participants were sent a check for the amount of their account balance.
In order to ensure adequate experimental conditions, a “shadow market” was established reflecting realistic wholesale costs to form the basis of price signals to participants. The experiment was conducted in three phases. During the summer, the wholesale supply feeder capacity for participants was virtually limited to 1,500-kW—this was large enough not to constrain the market. In the fall and winter, the feeder limit was reduced to 750-kW and 500-kW respectively. With those tighter constraints, demand-side management and distributed generation resources needed to be dispatched in order to balance the market.
Continuation of the work
Since the Olympic Peninsula Project was completed in 2007, there have been other similar demonstration projects. In 2010, the PNNL launched the Pacific Northwest Smart Grid Demonstration Project, a $173 million five-year project covering 60,000 metered customers in five states that aims to reduce peak loads, facilitate renewable resources, and improve system performance through a hierarchy of real-time price signals. In June 2014, the AEP Ohio gridSMART Demonstration Project published the results of its first phase. The gridSMART Project, which has strong similarities to the Olympic Peninsula Project, allows customers to bid their demand for energy, ancillary, and capacity payments.
Unless otherwise indicated, the following analysis will be based on data and findings from the PNNL’s Olympic Peninsula Project (2007).
Retail-level double-sided real-time electricity markets
The Olympic Peninsula Project was pioneering in that it was the first successful retail-level implementation of a double-sided auction market with five-minute clearing intervals.
In contrast to wholesale electricity markets, retail-level auctions are limited to the demand-side and distributed generation resources within a distribution feeder network. Supply from the bulk transmission is system is available within feeder constraints at the locational marginal price (LMP). If demand exceeds the feeder limit, the real-time price within the feeder increases until local supply and demand are at equilibrium.
The feature to note in a double-sided auction is that both the demand and the supply are being represented by their respective curves. This is important because resources that do not clear the auction remain inactive until they clear a subsequent auction. In the Olympic Peninsula Project, both supply-side and demand-side resources participated in the retail-level auction. On the supply side, parallelable (grid-interconnected) fossil-powered generators bid according to their startup cost and marginal cost of operation. On the demand side, responsive loads including HVAC units and municipal water pumps bid their value of lost load (VoLL). When a demand-side resource cleared the auction, it automatically curtailed its load until the next auction. Two backup generators also bid on the demand side, as their grid-isolated operation constrained them to reducing the load they were serving.
In the Olympic Peninsula Project, the retail-level double-sided auction cleared every five minutes. Each bidding interval, the clearing price was published and devices reacted accordingly. The green diamonds in Figure 1 show the five-minute retail-level local marginal prices during the period of October 30 to November 1, 2006. When the aggregate customer demand was below the 500-kW feeder limit, the energy price equaled the wholesale LMP. When the feeder’s capacity was exceeded, the price increased, loads were curtailed, and generators entered until supply and demand were at equilibrium.
Figure 1: The double-sided five-minute retail-level auction market
The interaction between supply and demand in retail-level double-sided auctions produces an interesting relationship between price and quantity. As is illustrated in Figure 2, when demand is very low the feeder itself is the marginal energy supplier and the clearing price is the wholesale LMP—as long as demand remains below the feeder capacity, quantity demanded fluctuates and price remains constant (a). Once demand reaches the feeder limit, demand-side bids become the marginal resource—because demand-side resources respond to price signals by curtailing loads, between the feeder limit and the lowest-cost distributed generator the price fluctuates and quantity demanded remains constant (b). At some point, the price rises to a sufficient level such that a supply-side resource clears the auction—when this happens, the cleared generation unit becomes the marginal producer, the price is fixed, and quantity fluctuates once again (c).
Figure 2: Understanding how real-time price can flatten load
Looking back at the three-day period from Figure 1, we can see in Figure 3 that there was indeed an alternating dynamic between quantity and price fluctuations. Until the total demand reached the feeder limit, the price was constant at the wholesale LMP. Once demand hit the feeder limit, controllable loads were curtailed and the price increased accordingly until a supply-side resource cleared the auction.
Figure 3: Control of imposed distribution constraint using price signals
Automation, algorithms, and the economics of load management
Much of the success of the Olympic Peninsula Project and other demonstration pilots was attributed to automation, rather than the actions of participants or operators. Participants tended to spend very little time directly managing their energy usage. In fact, 55 percent of final survey respondents could not recall whether they were assigned to a fixed, time-of-use, or real-time pricing arrangement.
In order to conduct double-sided retail-level auctions, bids need to be aggregated through an automated process. As is illustrated in Figure 4, there are three basic steps:
- Bids from price-responsive appliances are determined based on current usage needs and aggregated behind the meter into overall customer price flexibility curves
- The distribution system operator’s grid management software aggregates anonymized demand curves from all customers and inserts them into the double-sided auction
- The double-sided auction determines the clearing price that balances the market and broadcasts it back to customer meters, which in turn collect updated appliance bids
Figure 4: Market-level automation of demand-side transactions
Behind the meter, appliances such as HVAC units and water heaters respond to price signals based on their respective bids. The determination of bids by appliances is automated according to algorithms that consider historical prices, their desired state, and in some instances the forecast of expected price signals.
In order to simplify the decision-making process for occupants, the parameters for appliance automation can be integrated into intuitive comfort settings (see Table 1) with underlying operational values. To illustrate, we take a closer look at the bid and response strategy for thermostatically controlled loads such as HVAC units. As shown in the formulas below and illustrated in Figure 5, the occupant’s comfort setting and desired temperature determine a bidding curve with a minimum, a maximum, and an elasticity (slope). The resulting bid price is in function of the current temperature relative to the desired state and adaptively adjusts to the retail-level market by reflecting the historical price average and standard deviation.
HVAC bid into price-flexibility curve
Thermostat response to market clearing price
where Pbid is the bid price, Paverage is the mean price of electricity for the last 24-hour period, σ is the standard deviation of the electricity price for the same period, and kT and Tlimit are chosen from kT_L, kT_H and Tmin, Tmax, depending on where Tcurrent presently is on the bid curve.
Figure 5: Bid and response strategy for thermostat-controlled loads
Table 1: Thermostat comfort settings and resulting kT values
Once the bids are posted, the market can clear either internally (i.e. within a building) or externally, as was done in the Olympic Peninsula Project. After receiving the resulting market clearing price Pclear, the adjusted set point Tset,a is calculated. In the case of Figure 5, the clearing price is significantly below the bid and the AC unit increases its energy consumption.
As can be seen in Table 1, some of the comfort settings offer the option of incorporating pre-cooling or pre-heating by setting the temperature limit further than the thermostat setpoint. Because the occupant in Figure 5 has elected to enable pre-cooling, the relatively low clearing price results in a thermostat response that is below the desired temperature. By responding in this manner, the HVAC system is effectively storing energy that can be released through load curtailment during peak real-time price periods.
A snapshot summary of real-time pricing participants’ comfort setting choices is shown in Table 2. The default option assigned to all participants at the beginning of the experiment, Balance Comfort, was unsurprisingly the most common setting at the end as well.
Table 2: Summary of real-time participants’ comfort level choices
Comparing the results of time-of-use and real-time pricing participants
In the Olympic Peninsula Project, time-of-use rates were the only pricing arrangement to exhibit meaningful reductions overall energy consumption. This conservation benefit was in addition to off-peak savings. Although the variances in the results were large, the data showed that the difference was statistically significant with a 95 percent confidence interval.
Table 3: Overall energy savings by pricing type
Furthermore, time-of-use pricing was the most effective at reducing peaks for entire residential loads. Indeed, the only consistently measurable energy-use impact that could be observed was the energy-use reduction of time-of-use participants during peak hours. Although critical peak pricing was available, it was only called upon once.
Table 4: Time-of-use on-peak and off-peak rates by season
Time-of-use control, however, resulted in abrupt—not smooth—load shifts during the start and end of the peak intervals. In aggregate and at scale, the effect of these abrupt changes in demand could be detrimental to system stability. Furthermore, time-of-use participants at times exhibited improper assignment of the peak interval; for example, people get out of bed later during the weekend. This had the effect of exacerbating the peak rather than reducing it.
In contrast to time-of-use pricing, the behavior of real-time participants was the smoothest, reflected actual system conditions, and exhibited peak usage reduction at times when it was most needed. This may explain why real-time pricing arrangements did not produce the lowest average peaks. During times of high energy demand and feeder constraint, the large real-time price differential between peak and off-peak hours shifted the consumption of real-time pricing participants significantly more than time-of-use pricing.
HVAC loads served by real-time pricing arrangements effectively used energy in the very early morning hours when electricity was least expensive. This effect was more pronounced when the feeder was constrained, although it occurred during both constrained and unconstrained feeder conditions.
Figure 6: Shifting of HVAC load in response to real-time prices
Although the system had no explicit predictive capabilities, the thermostats of real-time pricing participants “learned” to avoid the mid-morning peak based on the diurnal shape of the price signal itself. The shift was more pronounced for occupants who chose to enable pre-heating or pre-cooling in their comfort settings. This result exceeded the project’s expectations.
Despite these interesting observations for some real-time pricing participants, the effect on aggregate demand was difficult to discern. Furthermore, severe weather conditions severely impaired the response to real-time price signals, which highlights the importance of having a diverse and substantial amount of price-sensitive load available in order to maintain system stability during such events.
Implementation and technical challenges
While there were some technological challenges during the course of the Olympic Peninsula Project, the research team found no fundamental limitations that would prevent the application of these technologies at a larger scale.
The first set of problems came during the recruiting phase. Many utility customers expressed early interest in energy conservation, however some had difficulty making the connection to load management and were not willing to give up absolute control of their thermostat or water heater.
Recruiting goals were not easily met and the signup period lasted longer than expected despite the financial incentives. Some applicants lacked fundamental knowledge about their appliances and Internet services, for example not knowing for certain whether they used electric or gas water heat. Despite the significant effort invested to recruit 200 participants, the project only located, qualified, and successfully signed up 112 homes.
The technical requirements to participate were as follows:
- HVAC space conditioning
- Broadband connection
- Meter within 60ft of house
- Electric water heater
A few common problems were experienced that disqualified applicants:
- Incompatible heating systems (no single thermostatic control)
- Inadequate Internet access (dial-up or intermittent)
- Physical location of the electric meter too far away from house
Once approved, each participant was provided the following equipment for the project:
- Home gateway
- Virtual Private Network (VPN) for those homes possessing digital subscriber line (DSL) broadband connections
- Water heater load-control module
- Communicating thermostat
- Advanced revenue meter
One these initial hurdles were cleared, overall customer complaints were minimal and the majority of initial participants remained with the experiment until its completion. According to the available information, the provided equipment seems to have performed as expected.
The Olympic Peninsula Project, which was conducted in 2006 and 2007, built its communications and control systems based on two types of telemetry:
- Broadband communication between home gateways and centralized systems
- Wireless telemetry of energy-management system data within residential premises
The reliability of telemetry was essential to the proper operation of the real-time market. Total passive (unresponsive) residential loads reported by participants’ meters also affected individual real-time bids for controllable loads. In order to maintain experimental conditions, meters that failed to report their load within 5 minutes before a market clearing were excluded from the total virtual feeder load for that market.
In Figure 7, it can be observed that two months were needed at the beginning to improve communications to a steady level that was then maintained until the end of the experiment. The reliability of communications ranged from 55 to 80 percent and differed among the different pricing groups—the bandwidth communicated by the equipment of each contract type was similar, so this difference cannot be easily assigned.
Figure 7: Network telemetry performance (15-minute intervals)
Despite a lack of reliability in telemetry, the system was built such that any loss of communication or failure in the market clearing signal resulted in devices falling into a safe default mode. Operations continued sub-optimally—but without catastrophic failure—until normal operation was restored. The research team concluded that this would scale well.
Stakeholder reactions and lessons learned
Customer feedback on the Olympic Peninsula Program was overwhelmingly positive:
- When the project end date was extended, most of the residential participants asked to continue—presumably because of the positive impact on their energy consumption
- Participants expressed a preference for and eagerly adopted time-of-use and real-time pricing arrangements over fixed prices
- After being in the project for a year, 73 percent of participants stated that they would choose a price-responsive arrangement in the future if given the opportunity
- During the closing survey, 95 percent of participants indicated that they would be likely or very likely to participate in a similar project in the future
The simplicity of participation, automation, and ability to override controls—which the research team refers to as the “friendliness” of participating in the demand response program—may be vitally important to achieving meaningful scale and magnitude of demand-side resources.
Consistent with this observation, the project provided all participants a means by which they could temporarily override price-responsive controls. In practice, however, very few participants appeared to have asserted their right to override appliance automation controls.
From the utilities’ perspective, the project was perceived as a success in that it assessed the feasibility of the technology in terms of “the efficacy of the interface between consumer and wholesale markets, the impact of consumer choice of service contracts in exchange for the opportunity to get paid to respond to system conditions, and the efficacy of GFA technology to manage under-frequency events”.
A few issues and concerns were identified from the utility perspective:
- The scale would need to reach several hundred MW in order to contribute significantly to frequency management and overall system reliability
- The ability to reprogram response may be essential because under some under-frequency events it may not be desirable for loads to react immediately
- The demand-side resources themselves are subject to outages when local power or Internet service become unavailable, which could be problematic in an emergency
- Much needs to be done in technician training, marketing and sales, standardization, financing, insurance, billing, customer relationship management, and other areas
Conclusions and experimental results
One of the primary goals of the Olympic Peninsula Project was to manage congestion on a distribution feeder. Based on the experimental results, this goal was accomplished. Using demand-side resources and dispatchable distributed generation, the virtual market designed by the project was able to operate the system within the feeder constraint for all but one five-minute interval during the course of the project year. As can be seen in Figure 8, distributed generation (the green line) and demand-side management (the gap between the green and blue lines) were able to maintain system reliability with a 500-kW feeder limit.
Figure 8: Duration of feeder capacity during 500-kW constraint
The capacity of demand-side resources and distributed generation was seamlessly offered and cleared through the project’s real-time retail-level double-sided auction market. As such, market-based control was shown to be a viable, effective tool for obtaining useful price-based responses for the entire feeder.
Conservative project estimates point to peak load reductions of 5 percent and 20 percent for 750-kW and 500-kW feeder constraints, respectively. Because the 1500-kW feeder limit was never reached, there was no peak reduction during that phase of the project. It should be noted that during constrained periods, the peak reductions for many individual weeks greatly exceeded the reported averages for the overall period.
Table 5: Average peak reduction during constrained periods
Comparing the performance of fixed, time-of-use, and real-time pricing arrangements, the project found interesting results. Although the overall energy conservation and peak reduction was highest for time-of-use pricing, real-time pricing participants demonstrated a more optimal shifting of load than for either fixed or time-of-use pricing participants.
Another interesting observation was that fixed-price participants sometimes shifted their load more than time-of-use participants. While this may be due to sampling limitations or coincidence, it does suggest than utility customers will respond manually to well-designed visual indicators and information sources about their consumption and cost.
The staff at PNNL conducted an intriguing analysis which, although not clearly conclusive, may be worth exploring. Applying the CAPM model to the mean and variance of peak energy usage for fixed, time-of-use, and real-time pricing participants, the research team at PNNL devised an “efficient frontier” that maximizes the combination of the three contract types. The main motivation in conducting this analysis seems to have been impact on utility revenues.
Figure 9: Efficient frontier between the three pricing arrangements
Results from the Ohio gridSMART project conducted between June and September 2013 were also promising. Wholesale energy purchases were reduced by 5%, a 6.5% peak load and 10% peak feeder load reduction were achieved at 50% simulated RTP household penetration, participants saw a 5% average reduction in household energy bill, and over 75% of participating customers were satisfied—40% of customers were very satisfied.
The PNNL’s five-year, 60,000 customer Smart Grid Demonstration is still underway and we are eagerly awaiting the results. The future looks bright for demand participation.