Gemini: A Non-Invasive, Energy-Harvesting True Power Meter

Bradford Campbell and Prabal Dutta

Electrical Engineering and Computer Science Department
University of Michigan
Ann Arbor, MI 48109
{bradjc,prabal}@umich.edu

Abstract—Power meters are critical for submetering loads in inside of a circuit panel dangerous and expensive. Two recent
residential and commercial settings, but high installation cost and designs, Monjolo [13] and a piezoelectromagnetic device [25],
complexity hamper their broader adoption. Recent approaches propose contact-less methods to power the sensors by using
address installation burdens by proposing non-invasive meters
that easily clip onto a wire, or stick onto a circuit breaker, energy-harvesting techniques, but both provide limited ac-
to perform contactless metering. Unfortunately, these designs curacy because without access to the voltage channel they
require regular maintenance (e.g. battery replacement) or reduce can only measure apparent power, at best, and certainly not
measurement accuracy (e.g. work poorly with non-unity power true power. A small, wireless, non-contact, easy-to-install,
factors). This paper presents Gemini, a new design point in and accurate power meter would overcome the limitations of
the power metering space. Gemini addresses the drawbacks of
prior approaches by decoupling and distributing the AC voltage current submetering devices.
and current measurement acquisitions, and recombining them To address these challenges, we present Gemini, a
wirelessly using a low-bandwidth approach, to offer non-invasive power metering system that decouples voltage measurement
real, reactive, and apparent power metering. Battery maintenance from current measurement to create a non-contact, energy-
is eliminated by using an energy-harvesting design that enables harvesting power meter capable of measuring true power. Two
the meter to power itself using a current transformer. Accuracy
is substantially improved over other non-invasive meters by insights enable this design point: 1) a power meter can harvest
virtualizing the voltage channel—effectively allowing the meter to energy to power itself from the output of a current transformer
calculate power as if it could directly measure voltage (since true that it also uses to measure the current waveform and 2) the
power requires sample-by-sample multiplication of current and AC voltage waveform can be virtualized by a single voltage
voltage measurements acquired with tight timing constraints). monitoring device [22], and key voltage parameters, instead
Collectively, these improvements result in a new design point that
meters resistive loads with 0.6 W average error and a range of of actual samples, can be used by the contact-less power
reactive and switching loads with 2.2 W average error—matching meters to compute true power. These insights enable a split
commercial, mains-powered solutions. metering architecture in which one or more plug-in devices
are responsible for measuring and virtualizing the AC voltage
I. I NTRODUCTION waveform and the majority of the other devices act as power
The need for low-cost, open-source, easily-deployable and meters that sample the local AC current waveform, locally
installable building-scale energy submetering devices is criti- synthesize the voltage signal, and calculate the power draw of
cally high, as noted by the U.S. National Science and Tech- their load (by multiplying voltage and current measurements
nology Council [19], the U.S. National Science Board [23], on a time-synchronized, sample-by-sample basis).
and the U.S. Department of Energy [14]. These reports exist This split architecture permits an energy-harvesting power
because buildings consume 73% of the electricity in the meter to accurately calculate real, reactive, and apparent power
U.S. [11] but “no widely-available tools exist to help tenants while remaining non-contact, which reduces size, cost, and
understand their energy consumption.” Submetering devices installation complexity. Gemini can replace a monolithic panel
that meet these requirements and can help address this problem level power meter that requires a control box and wires for
remain elusive as current designs are forced to trade-off every circuit, with small, split-core current transformer-sized
between installation ease, accuracy, cost, and maintenance. devices that can easily be clipped onto circuits of interest.
Single, self-contained devices [10], [16] require physical Gemini can also function as a compact plug-load meter to
contact to the AC mains in order to measure the AC voltage provide plug-load-level submetering. Both versions wirelessly
signal. This is acceptable for plug-load meters, but ill-suited transmit their results to the cloud over a low-power network.
for circuit level metering inside a panel box. Existing panel Two central concepts—energy-harvesting and voltage
metering solutions require a complex installation procedure virtualization—enable the Gemini design, but they also add
involving wiring to a spare breaker for capturing voltage and operational constraints not found in other meters. Energy-
running current transformer leads for each circuit to a central harvesting solves the issue of powering the meter, but it only
metering device [4]. Panel meters effectively measure at the provides enough energy to intermittently power and activate
circuit level but cannot provide insight on a per-load basis. the device for very short periods of time. To operate under
Additionally, these power meters typically use AC mains and power intermittency, we partition the power measurement task
a high voltage AC-DC power supply to power themselves, across multiple activations to provide sufficient energy for
but requiring a connection to the AC mains makes installation communications, sensing, storage, and computation.
 Type Is Harvests Calculates
Meanwhile, voltage virtualization requires a shared notion Wireless Energy True Power
of time between the device parametrizing the voltage and the ACme [16] Plug-load Yes No Yes
device measuring the current, to properly align the phase of the Watts Up? .net [10] Plug-load No No Yes
WeMo Insight [1] Plug-load Yes No Yes
two signals. Prior work uses clock synchronization protocols,
TED [8] Split-core No No Yes
but such approaches are incompatible with the intermittency Magnetometer [20] Magnetometer Yes No Yes
imposed by energy-harvesting operation. To address this prob- Virtualization [22] Current transformer Yes No Yes
lem, we transfer relative time (or phase offset) using a single Stick-on [25] Piezoelectromagnetic Yes Yes No
Monjolo [13] Plug-load, split-core Yes Yes No
packet transmission, allowing the voltage sensor to propagate
Gemini Plug-load, split-core Yes Yes Yes
its local time across domains to the current sensor/power
meter’s local time. TABLE I: Comparison with prior work. Gemini is unique in
To evaluate the architecture, we prototype both the voltage that it is wireless, harvests energy, and calculates true power.
monitoring device and the energy-harvesting power meter.
We show that virtualizing the voltage channel is a viable
approach to designing a power meter. We evaluate the system C. Non-Contact Meters
by benchmarking its performance with resistive, inductive, and Realizing that contacting the AC line for power metering in-
switching loads, including ones with dynamic power factors. curs high installation costs and limited deployment scenarios,
We find that Gemini can meter resistive loads with an average many recent meters have explored various non-contact options.
error of 0.61 W and inductive and switching loads with an 1) Non Energy-Harvesting: Patel et al. designed an exter-
average error of 2.2 W. We also demonstrate that the energy- nally powered house-level stick-on power meter using magne-
harvesting power meters can harvest and operate with loads tometers designed to be attached to the circuit breaker in the
as low as 0.5 W. Further, we evaluate the effect of timing in house [20]. The sensor measures only current, which is later
the system and show how timing errors lead to phase offsets multiplied by RMS voltage by a PC. This split approach is
and measurement error. Finally, we discuss a technique for architecturally similar to Gemini but uses RMS values instead
matching the correct virtualized voltage channel to the current of true readings which cause significant error for reactive and
channel under measurement, describe the trade-offs inherent switching loads. Also, while unobtrusive for a single sensor,
in the design space, and outline avenues for future work. monitoring individual circuits and powering the sensors to
II. R ELATED W ORK do so is challenging and error-prone due to crosstalk from
adjacent circuits.
A range of commercial and research power metering sys- An approach proposed by Schmid et al. is similar to Gemini
tems have been designed. Table I compares power meters in that it uses one sensor to parametrize and transmit the
across three key dimensions: the method used to transmit voltage channel to another time-synchronized node that is able
measurements, the method used to power the meters, and to synthesize the voltage waveform [22]. The final step of
whether the meters can calculate true power (i.e. whether they computing power is done offline. Gemini extends this idea by
calculate power by multiplying voltage and current measure- synthesizing voltage, measuring current, and computing power
ments acquired in a synchronized manner on a sample-by- on a single node, by powering that node by energy-harvesting
sample basis). Only Gemini achieves all three. from a current transformer, and by removing the need for a
A. Plug-Load Meters wireless time-synchronization protocol which is challenging
Plug-load meters [1], [3], [7], [10], [16], [17], [21], [24] to maintain with the limited operating energy budget provided
are capable of providing high-fidelity, accurate power mea- by energy-harvesting.
surements for single loads. However, for loads that are hard 2) Energy-Harvesting: A piezoelectromagnetic (PEM) de-
to move (e.g. kitchen appliances) or built in (e.g. lighting), vice designed by Xu et al. is both an energy-harvesting
attaching a plug load meter is difficult or impossible. Also, device that can power a sensor node and a current sensing
plug-load meters are typically active, meaning that they draw device for circuit-level metering [25]. The PEM is a “stick-
power even when the metered load is off. In contrast, Gemini on” device that attaches to each breaker in a circuit panel
supports plug-load or circuit panel installation, and offers to meter current. However, the PEM suffers from crosstalk
power-proportional power metering (i.e. zero idle power). and, without access to the voltage channel, cannot provide
true power measurements.
B. Whole-House Meters The Monjolo [13] energy meter combines energy-harvesting
An alternative to plug-load metering is whole-house meter- using a current transformer and energy metering. The device
ing in which a single meter monitors the power draw of the harvests at a rate proportional to the the power of the attached
entire building. Devices like the The Energy Detective [8] and load and converts that rate to a power measurement. This
Blue Line Innovations’ PowerCost Monitor [6] provide this indirect sensing works for fully resistive loads but exhibits
data but have high installation overhead and are not capable inaccuracies for loads with non-unity power factors.
of submetering. Gemini can also meter whole-house power by In contrast to these designs, Gemini uses an energy-
attaching to a home’s drip-loop or incoming AC line, but it harvesting node to sample the current waveform, wirelessly
additionally supports load and circuit level submetering. acquire the voltage parameters, and compute true power.
 200
AC
V
150
100

Voltage (V)
Load A
VAC 50
0
IAC Receiver -50
2
-100
EH V? -150
VAC<A,φ> Power -200
3 0 5 10 15 20 25 30 35 40 45 50
Load B 1 Time (ms)
Cloud
Voltage (A) Voltage (B)
IAC
EH
EH IAC Fig. 3: Comparison of two voltage waveforms at physically
Load C
disparate points in the same electrical system. Measuring the
voltage at two different loads with different power draws on
different circuits tapped off of the same leg of a transformer
Fig. 1: Gemini overview. A single voltage monitor (V) con-
results in voltage waveforms that are nearly identical. This
tinuously monitors the voltage channel of the AC circuit.
suggests that measuring voltage at a single watchpoint is
Energy-harvesting power meters (EH) use current transformers
sufficient for calculating the power draw of multiple loads.
to meter each load. To calculate true power, the EH meter
requests the voltage parameters from the voltage monitor,
samples the local current waveform, synthesizes the voltage That is, voltage can be modeled with phase information and
waveform, computes power, and transmits the result to a one or a few Fourier coefficients. Figure 2 provides some
receiver which forwards the data to the cloud. Maintaining intuition for this distinction by showing the voltage and
timing across domains is key to the success of this approach. current waveforms for several loads. The voltage waveform
is consistently sinusoidal while the current exhibits significant
harmonic distortion. Because sending raw waveform samples
III. S YSTEM OVERVIEW
is infeasible due to limited packet payloads and energy re-
Typical AC power meters simultaneously sample the voltage sources, virtualizing the voltage channel and transmitting its
and current channels and multiply these signals on a sample- smaller parameter set is the preferred approach.
by-sample basis to calculate power. Gemini, in contrast, is an Another tenet of the Gemini design is that only a small
AC power metering system in which the voltage and current number of voltage monitoring sensors are required to support
waveforms are measured on separate devices and wirelessly a much larger number of current sensing power meters. This
recombined to calculate power. This “split” architecture allows holds if the local voltage measurement can be replaced with
for a non-contact, energy-harvesting, true power meter. a voltage measurement obtained elsewhere in the load tree.
Figure 1 shows the two Gemini components, the voltage To verify this, we sample the voltage waveform at two points
monitor (V) and the energy-harvesting power meter (EH), on the same leg of a transformer from two different circuits
which work in tandem to measure power. The voltage monitor behind two different electrical panels with a 300 W load on
continuously monitors the voltage channel and provides a one point and a 1.5 kW load on the other. The resulting voltage
virtualized representation of the voltage waveform to the waveforms are shown in Figure 3. The waveforms show no
energy-harvesting nodes on-demand. Each energy-harvesting phase distortion and a slight amplitude variation, suggesting
node measures its local current waveform and multiplies that that remotely sensing voltage is viable. One other important
by the voltage waveform obtained from the voltage monitor factor related to voltage measurement is that the phase of
to calculate true power. The energy-harvesting node then the voltage channel local to the power meter (which is not
transmits the power data to the cloud via a receiver. measured) must match the phase of the virtualized voltage
channel used for power calculation. A method for matching
IV. D ESIGN channels is discussed in Section VII-B.
Gemini’s design employs distributed voltage and energy-
harvesting current monitors cooperatively calculating power. B. Energy-Harvesting
To meet the ease-of-installation constraint, the current sen-
A. Virtualizing Voltage sor must be able to sample the current waveform while
To realize this split metering architecture, and to success- remaining non-contact. To power the sensor, we are unable to
fully calculate power with Gemini, either the current mea- use an AC mains based AC-DC converter because that requires
suring device must learn about the voltage waveform from tapping into the circuit, and we do not use batteries because
the voltage monitor or vice-versa. In our design, we transmit of their size and the maintenance burden of replacing them.
information about the voltage waveform to the current sensor Instead, we use an energy-harvesting power supply that uses
to calculate power. The reason for this is that the voltage the output of a current transformer to charge a capacitor. This
channel is more easily parametrized than the current channel. approach is largely borrowed from the Monjolo meter [13].
 200 2 200 2 200 2 200 2
150 1.5 150 1.5 150 1.5 150 1.5
100 1 100 1 100 1 100 1
Voltage (V)

Voltage (V)

Voltage (V)

Voltage (V)
Current (A)

Current (A)

Current (A)

Current (A)
50 0.5 50 0.5 50 0.5 50 0.5
0 0 0 0 0 0 0 0
-50 -0.5 -50 -0.5 -50 -0.5 -50 -0.5
-100 -1 -100 -1 -100 -1 -100 -1
-150 -1.5 -150 -1.5 -150 -1.5 -150 -1.5
-200 -2 -200 -2 -200 -2 -200 -2
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Time (ms) Time (ms) Time (ms) Time (ms)
Voltage Current Voltage Current Voltage Current Voltage Current

(a) 40 W Incandescent (b) AC Fan (c) 3D Printer (d) Refrigerator

200 2 200 2 200 3 200 2
150 1.5 150 1.5 150 2.25 150 1.5
100 1 100 1 100 1.5 100 1
Voltage (V)

Voltage (V)

Voltage (V)

Voltage (V)
Current (A)

Current (A)

Current (A)

Current (A)
50 0.5 50 0.5 50 0.75 50 0.5
0 0 0 0 0 0 0 0
-50 -0.5 -50 -0.5 -50 -0.75 -50 -0.5
-100 -1 -100 -1 -100 -1.5 -100 -1
-150 -1.5 -150 -1.5 -150 -2.25 -150 -1.5
-200 -2 -200 -2 -200 -3 -200 -2
0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50 0 5 10 15 20 25 30 35 40 45 50
Time (ms) Time (ms) Time (ms) Time (ms)
Voltage Current Voltage Current Voltage Current Voltage Current

(e) MacBook Pro (f) Dimmed Light Bulb (g) Ice Shaver (h) Audio Receiver
Fig. 2: Voltage and current waveforms of eight loads. The current waveforms vary wildly, but the voltage waveforms are
consistently sinusoidal and can be represented by one or just a few Fourier coefficients. The sinusoidal nature of the voltage
waveform makes it a better candidate to be virtualized, thus supporting our architectural choice to virtualize voltage.

200
Current (A) Voltage (V) Voltage (V)

100
0
-100
-200
200
Synthesized

100
0
-100
-200
1
0.5
0 (a) Voltage Monitor (b) Energy-Harvesting Node
-0.5
-1 Fig. 5: Sensing hardware.
1
Current (A)
Sampled

0.5
0
-0.5
-1 C. Power Meter System Operation
-10 -5 0 5 10 15 20 25 30
Time (ms)
Power is calculated by the energy-harvesting current sensor.
Fig. 4: Voltage and current waveforms. Ground truth cur- The sensor begins by requesting information about the voltage
rent and voltage waveforms are shown above their synthe- channel from an always-on voltage monitor. After sending
sized/sampled counterparts. The voltage waveform is synthe- the request, it samples the current waveform until the voltage
sized from the parameters sent by the voltage monitor and monitor responds. It then synthesizes a discrete representation
the current waveform is sampled from the current transformer of the voltage channel properly phase-aligned to the sampled
monitoring the load. The discrete waveforms are multiplied current waveform, as shown in Figure 4. It multiplies the
together, point-by-point, to compute the load’s power draw. voltage and current samples together and averages them to find
the instantaneous power draw of the load. Harvested energy
permitting, the current sensor sends the power measurement
This power supply is necessarily intermittent; harvesting to a receiver which timestamps and forwards the data to the
from a current transformer does not supply sufficient power cloud, otherwise the sample is stored for future transmission.
to run a sensor node continuously. It is also unpredictable, as The receiver is an always-on listener that may be a dedicated
variable-power loads will alter the operation of the harvester. gateway for the low-power wireless network or it may be a
This causes the power supply to limit the sample rate of the mains-powered node, such as a voltage monitor, which can
meter based on how quickly and often it can recharge. forward the packet on behalf of the meter.
 3.3V
Voltage Waveform
160Ω 100kΩ

0.1μF 100kΩ

Current Waveform
ADC

AC Phase
Power VCC Core
Supply Timing
t0
57μF (MSP430+

SFD
Voltage Monitor A1..n, t0, f
600μF (LTC3588) CC2420)

TX TX

SFD
Energy Harvester V? pkt A1..n, t0, f
Fig. 7: Energy-harvesting power supply and measurement
circuit. One current transformer is used for harvesting and
Request Sample Receive t
Voltage
Parameters
Current
Waveform
Voltage
Parameters
the other for measurement. Future implementations could
multiplex one current transformer for both operations.
Fig. 6: Transmitting AC voltage phase information. After the
energy-harvesting node requests the voltage parameters, the
voltage monitor responds with parameters of the voltage wave- 6
VCAP
form such as Fourier coefficients, phase, and frequency. Phase 5
is encoded as the time offset (t0 ) between the response packet’s
4

Voltage (V)
SFD signal and the voltage signal’s most recently rising zero- VCC
crossing. The energy-harvesting node uses its local timestamp 3
of the SFD signal to phase align the current waveform samples Sample Current
2 Application
to the voltage channel without requiring synchronized clocks. Starts Receive
Voltage
1 Request Write FRAM
Voltage
V. I MPLEMENTATION 0
0 50 100 150 200 250 300 350
To evaluate our hypothesis and explore the performance of Time (ms)
our design, we implement a prototype version of Gemini.
Fig. 8: Energy-harvesting operation. VCAP is the voltage on
A. Voltage Monitor the 600 µF capacitor and VCC is the 3.3 V supply voltage.
The voltage monitor continuously records and can on-
demand report parameters of its local AC voltage signal.
Our implementation uses a TI CC2538 SoC [9] for compu- The receiver uses its local SFD timestamp and the rising
tation and wireless communication, and an Analog Devices zero-crossing time offset provided by the voltage monitor to
ADE7753 [12] for AC voltage waveform analysis. align the voltage and current waveforms. Any delays in packet
The voltage monitor stores the peak voltage of the most transmission and processing are ignored because all offsets are
recent AC cycle and locally timestamps the most recent rising calculated from the synchronized SFD signal.
zero-crossing of the voltage waveform. This allows the voltage
B. Energy-Harvesting Power Meter
monitor to make a simple parametrization with the phase
and magnitude of the voltage waveform for dissemination to The energy-harvesting device (Figure 7) is responsible for
interested clients. The ADE7753 also provides access to digital sampling the AC current waveform and calculating the instan-
samples of the entire voltage waveform allowing for more taneous power draw of the load or circuit.
sophisticated parametrizing schemes to be used in the future. 1) Power Supply and Coil Circuit: The energy-harvesting
The voltage monitor provides the parametrized voltage power supply is based on the Monjolo power meter’s de-
information as a service. An interested client sends a wireless sign [13]. The harvesting source is the AC output of a current
packet to the voltage monitor requesting the parametrized transformer that is wrapped with 10 windings of the AC
values. The monitor immediately replies with the magnitude phase line. This AC signal is fed to a Linear Technology
and phase of the voltage signal. Because we do not assume LTC3588 [18]. The LTC3588 is configured to charge a 600 µF
a synchronized timebase between the client and monitor, the capacitor bank to 5.25 V and then enable a 3.3 V output
phase information is transmitted as a time offset between the regulator that powers the main system. When the main storage
start of frame delimiter (SFD) signal of the outgoing reply capacitors discharge to below 3.8 V, the regulator is disabled,
packet and the most recent rising zero-crossing of the voltage the storage capacitors begin recharging, and the system runs on
waveform, as shown in Figure 6. The transmitter and receiver the charge stored in the 57 µF output capacitor until the voltage
establish a common time reference by assuming the SFD drops too low and the system is forced off. One iteration of
signal asserts simultaneously on both devices. this cycle is shown in Figure 8.
 Activated
operating. This rate-limiting prevents a node from saturating
the wireless channel by transmitting packets too frequently.
Power result We size the timing resistor and capacitor to limit power
No stored calculated
Read
power result
FRAM last activation calculations to no more than once every five seconds.
Sample
Timing Transmit When an energy-harvesting node determines that it should
Capacitor Power
VTIME >
VTIME < Result calculate power, it requests the voltage parameters from a
threshold voltage monitor. After transmitting the request, it begins
threshold
Request Sample Receive sampling the current waveform every 40 µs while awaiting
Voltage Current Voltage
Params Waveform Params a response from the voltage monitor. The response from
the voltage monitor arrives approximately 16 milliseconds
Store Synthesize after the request, which allows the energy-harvesting node to
Calculate Voltage
Result in sample a full period of the current waveform. Upon packet
Power Waveform
FRAM
reception, the energy-harvesting node ceases sampling and
begins calculating the average power over the prior AC period.
Done The response packet from the voltage monitor contains the
offset from the SFD signal of the response to the last rising
Fig. 9: State machine for the energy-harvesting power meter. zero-crossing of the voltage waveform in terms of the local
Nodes use the timing capacitor to rate-limit sampling and clock on the voltage monitor. To convert that offset to the
determine when to compute power. local clock of the energy-harvesting node, the node computes
the number of current waveform samples since the voltage
zero-crossing. It then uses a precomputed array of sine wave
In our implementation, a second current transformer is used samples to multiply the current waveform samples with the
for sampling the current waveform. The output of this current synthetic voltage waveform to obtain the power waveform.
transformer is connected across a burden resistor and fed The power waveform samples are then averaged to calculate
first into a biasing circuit and then into an ADC (shown in the average power draw. Due to the limited energy budget from
the top of Figure 7). The burden resistor affects the scaling the energy-harvesting power supply, the power calculation is
factor between the actual current waveform and the current stored in FRAM and not transmitted immediately. On the next
transformer output and increasing the burden resistor increases activation the result is transmitted to the cloud. This imposes
the peak-to-peak output voltage. To ensure sufficient resolution a small time delay for the power measurement but allows the
of the voltage signal from small primary loads we use a system to operate within its energy budget.
160 Ω resistor. As we will show in Section VI, selecting
a smaller burden resistor is necessary for larger loads as VI. E VALUATION
the ADC saturates. The biasing circuit is required to shift We now evaluate the viability of Gemini as a power meter-
the output of the current transformer from being centered ing system, the effects of using an energy-harvesting power
around 0 V to being centered around 1.65 V. This ensures supply on Gemini’s operation, and how design decisions and
that all output waveforms up to 3.3 V peak-to-peak can be timing errors affect system operation.
successfully sampled by the ADC.
2) Computation Core: The energy-harvesting node is based A. Power Metering
on the Epic sensor node module [15] which provides a To evaluate Gemini’s performance as a power meter, we
microprocessor and radio. Our design adds a 4 Kb Cypress sweep over a range of resistive loads (i.e. loads with unity
FM25L04B FRAM module [2] for persistent storage between power factor) and measure how well Gemini tracks ground
activations, and a parallel resistor-capacitor circuit for roughly truth (as measured with a PLM-1LP [5]), as shown in Fig-
timing the period of activations to rate-limit activations for ure 10. For resistive loads below 5 W the energy-harvesting
high-power loads, as in the Monjolo design [13]. meter is unable to accumulate enough energy to activate.
3) Operation: The energy-harvesting node’s operation fol- Between 5 and 95 W, the estimated power tracks ground truth
lows the state machine shown in Figure 9. Upon activation, well with an average absolute error of 0.61 W and average
it reads the FRAM to determine if a power measurement percent error of 1.2%. Above 95 W the peak-to-peak output of
was calculated on the previous activation and needs to be the current transformer exceeds the 3.3 V ADC reference and
transmitted. If so, it transmits the power value to a receiver, clips, causing greater error. Reducing the value of the burden
transitions to the done state, and waits for the power supply to resistor would allow for these larger loads to be metered.
exhaust. If there is no stored power result, the node samples the While Gemini performs well for purely resistive loads (such
voltage of the timing capacitor to determine the approximate as incandescent light bulbs), many loads do not have a unity
elapsed time since the last activation. If the capacitor voltage power factor and a purely sinusoidal current waveform. To
is above a threshold, the node simply sits idle until the power evaluate Gemini with such loads, we meter four AC loads with
supply capacitors discharge and the node is forced off again. both Gemini and a ground truth power meter. The experimental
If the voltage is below the threshold, the node continues traces are displayed in Figure 11.
 80 100 90 100 9 100 150 100
70 80 8
80 70 80 7 80 120 80
60
60 6
Power (W)

Power (W)

Power (W)

Power (W)
50 60 60 60 90 60
50 5

%
40
40 4
30 40 40 40 60 40
30 3
20 20 2
20 20 20 30 20
10 10 1
0 0 0 0 0 0 0 0
0 5 10 15 20 25 30 0 10 20 30 40 50 60 0 10 20 30 40 50 0 10 20 30 40 50 60 70 80
Time (m) Time (m) Time (m) Time (m)
% Error PF % Error PF % Error PF % Error PF
Power Estimated Power Estimated Power Estimated Power Estimated

(a) 3D Print Job (b) LCD Monitor (c) Smartphone Charging (d) Desktop
Fig. 11: Power metering accuracy for four real-world loads. We meter four loads with a ground truth meter and Gemini. Gemini
is able to mirror ground truth well with the highest error appearing when loads either rapidly change power draw or when
loads draw very little power (e.g. drawing ~1 W). For instance, because (c) has a low power draw, the average percent error
is relatively high at 17.2%, but the average absolute error is only 1.0 W. The highest average absolute error is 4.9 W for (d),
mostly due to the current waveform clipping after the 55 minute mark. These test cases demonstrate Gemini’s viability as a
power meter for real-world loads that operate within Gemini’s (configurable) operating range.

7 in Gemini’s percent error as the energy-harvesting power

6 supply limits the sample rate of Gemini causing spikes to be
Error (W)
Absolute

5 Average Error: 0.61 W

4 reported late or missed all together. Overall, however, Gemini
3
2 reports an average percent error of 4.9% and an average
1
0 absolute error of 2.2 W.
20
Average % Error: 1.2%
15 2) 24” LCD Monitor: The monitor is cycled between the
Error (%)
Percent

10 on, off, on, and standby modes. During the on mode, Gemini
5 is quite accurate with an average error of 0.85% and average
0 absolute error of 0.67 W. In the standby and off modes Gemini
100
overestimates, but the overall total average error is still only
Estimated Power (W)

90
80 9.7% and 0.68 W. A key result of this test is that Gemini
70
60 continues to operate even when the monitor is drawing only
50
40 0.5 W in the low power states.
30
20
10
3) Charging Smartphone: In this test the power factor is
0 roughly constant at 0.58. Again Gemini tracks the slowly
0 10 20 30 40 50 60 70 80 90 100
Ground Truth Power (W)
decreasing power draw well. The average percent error is
relatively high at 17.2%, but because phone chargers draw
Fig. 10: Accuracy over a range of resistive loads. We meter little power, the overall average absolute error is only 1.0 W.
a range of loads with a unity power factor using Gemini. 4) Desktop Computer: The final load is a desktop computer
The x-axis shows the ground truth power draw of the load. that starts in sleep mode, enters normal operation for 55
In the bottom graph, Gemini’s estimate is overlaid on the minutes and then plays video for 20 minutes. At the beginning,
line representing an ideal device. The middle graph shows the energy-harvesting power meter is able to activate even as
the percent error of the estimate, and the top graph shows the the load draws only 1.1 W. During normal operation, Gemini
absolute error of the estimate. The average percent error is performs well, albeit with the same spikes in error as the load
1.2% and the average absolute error is 0.61 W. The error grows power changes rapidly like in the 3D printer case. In the final
when the load is greater than 95 W due to ADC clipping of phase of the test, when the desktop plays a video, Gemini
the current waveform. These results demonstrate that Gemini underestimates slightly. This is, as in the fully resistive case,
is a viable power meter for resistive loads. due to clipping of the current waveform when sampling it with
the ADC. Even with the clipping error, however, the overall
average error is 6.3% and 4.9 W.
1) 3D Printer: The printer starts in an idle state before For a rough comparison of Gemini’s error to commercially
heating the extruder head between minutes 4 and 6, then available meters, we compare the reported specifications of
actively prints before returning to idle at minute 24. Gemini several of them with the empirical data collected using Gemini.
tracks these transitions well, even being able to sample the The Watts Up? .Net [10] claims a 1.5% error, the Kill-A-
increasing power transition between idle and heating. Rapid Watt [3] reports a 0.2% error, and the PLM-1LP [5], our
swings in the power draw between 50 and 70 W cause spikes ground-truth meter, cites a 0.5% error.
Frequency (Activations/second)

(Activations/second)
2 1
Solid-core CT
Split-Core CT 0.8

Frequency
1.5 0.6
0.4
1
0.2
0.5 0
0 1 2 3 4 5 6 7 8 9 10 11
Coil Windings (#)
0
0 50 100 150 200 Fig. 13: Effect of the number of coil windings on system
Load Power (W) performance. The activation rate for the energy-harvesting
Fig. 12: Activations/second for two coils over a range of node with a 75 W load attached is shown as the winding count
resistive loads. The first coil is a solid-core current transformer changes. As the windings increase, the activation rate, and
(CT) with 10 windings of the AC phase line around it. The therefore the sample rate, increases. The increased windings
second is a split-core CT. Below 5 W for the solid-core and also increase the peak-to-peak voltage output of the current
below 117 W for the split-core the energy-harvesting node transformer, changing the range of loads that can be success-
is unable to harvest enough energy to activate and therefore fully measured with a fixed burden resistor.
is unable to perform any metering. Beyond that, as the load
power increases, the ability for the energy-harvesting node
to harvest increases. For comparison, a 100 W load causes To further examine the relationship between windings and
the solid-core version to activate at 1 Hz, and because the activation rate we fix the load power at 75 W and adjust the
power measurement is transmitted on the second activation, windings around the coil to observe how the activation rate
Gemini offers a 0.5 Hz sample rate. These sample rates are changes. Figure 13 shows the results. As expected, increasing
comparable with other power meters, demonstrating that an the windings increases the activation and sample rates. While
energy-harvesting power supply is a viable option. this aids the temporal fidelity of the meter, it causes the peak-
to-peak output voltage of the current transformer to increase,
reducing the range of loads that the energy-harvesting meter
B. Energy-Harvesting can measure for a given burden resistor. We discuss this, and
A critical element of the Gemini design is the energy- other trade-offs, further in Section VII-D.
harvesting power supply. A consequence of using energy- One important note to consider is that the data points
harvesting is that traditional, continuous metering is no longer relating to activation rate with different coils and windings
possible. Instead, power measurements are intermittent with are collected with a purely resistive primary load. In our other
a rate that depends on the power draw of the metered load. experiments from Section VI-A, the minimum power required
To quantify how often an energy-harvesting power meter for activation was much lower than 5 W. Based on this, the
can perform a measurement, we sweep over a range of performance of the energy-harvesting power supply depends
power draws using a fully resistive test load and monitor not only on the average current the load is drawing, but also
the activation rate of the sensor. We do this both with our on the shape of the current waveform. Therefore, the sample
main implementation that harvests from a solid-core current rate will vary based on the power factor of the load being
transformer wrapped with ten windings of the AC phase line metered.
and with a nearly identical implementation that uses a split-
core current transformer around a single wire. The results are C. Impact of Timing Errors on Power Measurement Quality
shown in Figure 12. For the solid-core version, the energy-
harvesting node is able to activate once the load reaches 5 W, The Gemini design hinges on the accurate transfer of timing
and it activates more rapidly as the load power increases. At information about the phase of the voltage waveform from the
100 W the sensor activates roughly once per second, which voltage monitor to the energy-harvesting meter. Timing errors
because the sensor needs two activations to take and transmit due to clock offsets, clock drift, or queuing delays cause phase
a measurement, equates to a maximum sample rate of 0.5 Hz. errors between the synthesized voltage and sampled current
For the split-core version, fewer windings mean a higher waveforms on the energy-harvesting meter, resulting in power
power load is required for the node to begin activating. At calculation errors. To characterize the effect of such errors,
approximately 117 W, the power supply is able to harvest we introduce artificial timing delays by shifting the voltage
sufficient energy to begin sampling. As in the solid-core waveform when calculating the average power of the eight
design, the sample rate increases with load power. Because loads in Figure 2. Figure 14 shows power estimation error as a
the split-core design clips around a wire it is most suitable function of the timing error. In the extreme, with ±180° phase
for metering entire circuits. At the circuit level, the power of error, or 8.33 ms timing error, we observe at least 200% error
multiple loads will sum and the higher minimum measurement in the power calculation. This demonstrates that minimizing
power will have less effect than with individual loads. timing error is critical to Gemini’s operation.
 Phase Offset (degrees) Calculated Power (W) for V-I Phase Error
Load
0 45 90 135 180 225 270 315 360 0° ±180° +120° +240°
250
AC Fan 11.777 -13.286 -14.932 0.588
Power Error (%)

200 Dimmed Bulb 31.991 -33.184 -22.175 -13.406

150 Light Bulb 40.564 -40.087 -14.745 -28.346
100 Audio Receiver 49.307 -51.782 -41.503 -14.238
50 3D Printer 50.926 -49.108 -11.304 -41.560
0 MacBook pro 65.880 -64.205 -17.888 -51.163
0 2 4 6 8 10 12 14 16 18 Refrigerator 102.382 -113.884 -131.309 9.436
Timing Error (ms) Ice Shaver 104.702 -107.55 -71.107 -44.028

40 W Bulb MacBook Pro TABLE II: Power calculations with different phase errors.
AC Fan Dimmed Bulb The instantaneous power for the eight loads in Figure 2 is
3D Printer Ice Shaver
Fridge Receiver calculated with the current and voltage waveforms correctly
aligned and with the current waveform shifted as would
Fig. 14: Effect of timing offsets when calculating power. Even happen if the incorrect voltage leg of a transformer was used.
1 ms of phase error can result in 50% measurement error. In all eight cases the correct result is the greatest positive
result, suggesting that using max() as a heuristic is often
sufficient for correctly selecting the voltage channel.
VII. D ISCUSSION
In this section, we discuss some limitations of the current
design, explore possible workarounds, account for sources of phases with the loads sampled in Figure 2. The results are in
error, and describe inherent design trade-offs. Table II. The second column shows the correct result with no
phase error. The remaining columns show the calculated power
A. Energy-Harvesting Limitations
result after subjecting the current waveforms to phase offsets
While energy-harvesting addresses issues with how to power that would occur when using an incorrect voltage channel.
the power meter, it has limitations. First, if the metered load For this small sample size, we find the correct phase yields
is turned off and does not draw any power, the meter does not the largest positive value, even for highly inductive loads like
activate or take measurements. However, a server can interpret the ice shaver and refrigerator.
the lack of activity as 0 W.
Second, if a load changes power state from active to off (as C. Sources of Phase Error
a refrigerator might do) between Gemini performing a power Due to Gemini’s distributed nature, several possible sources
measurement and it being able to transmit that measurement, of error could arise when aligning the phases of the current and
the reading could be severely time-shifted from when the load voltage waveforms. First, the ADE7753 introduces an average
was actually drawing that power. While this could represent offset of 23.2° between the rising zero-crossing and when its
a significant error, it is transient and quickly corrected by the zero-crossing interrupt line is asserted. We account for this
next measurement. average but additional jitter can cause phase errors. Second,
current transformers exhibit phase error offsets and the split-
B. Matching Current and Voltage
core version used in Figure 12 adds an average error of 1.5°,
Buildings typically have two- or three-phase power systems, for which we must account. Third, quantization errors that
meaning circuits on different legs of the transformer will have manifest from the discrete samples of the current waveform
voltage phase offsets between them. In order to correctly not perfectly aligning with the stored samples of the voltage
calculate power, Gemini must match the correct voltage phase wave can cause up to 40 µs or 0.86° of phase error.
to the load being metered. Loads with a highly displacement
based power factor could make selecting the correct voltage D. Design Space Trade-offs
channel ambiguous. In practice, however, two factors make Gemini operates in a rich design space with many axes on
this type of error unlikely. First, many loads have switching which to trade-off different goals. Our implementation chooses
power supplies which exhibit harmonic distortion to their one particular design point but helps illustrate how different
current waveforms but little displacement. Second, higher choices would affect the system. Key properties, like the range
power loads are required to include power factor correction of measurable loads, the rate of reporting, and measurement
which manipulates the current waveform to appear more like accuracy can all be adjusted. The range of loads that can be
a resistive load. measured is a function of the number of windings around
One way we could match the correct voltage channel is the current transformer and the burden resistor. Decreasing
to use a simple heuristic: each power meter computes power the windings or burden resistance increases the maximum
using all available voltage phases and selects the voltage phase load that can be measured. However, decreasing the windings
that yields the largest positive power value. Demonstrating this reduces the energy-harvester’s ability to harvest, reducing
heuristic, we simulate the effect of using incorrect voltage sampling fidelity.
 The sampling rate can be increased by requesting the R EFERENCES
voltage parameters less often and using the energy spent [1] Belkin WeMo Insight.
on the radio to keep a clock running. This would allow http://www.belkin.com/us/p/P-F7C029/.
[2] Cypress FM25L04B-DG.
a single voltage channel update to calculate several current http://www.cypress.com/?mpn=FM25L04B-DG.
measurements, at the expense of accuracy due to clock drift. [3] Kill-A-Watt Wireless. http:
Tuning these properties allows Gemini to be adapted for //www.p3international.com/products/consumer/p4220.html.
[4] Obvious Power Panel+.
applications with specific requirements. http://www.obvius.com/Products/PPP-O-XX.
E. Future Work [5] PLM-1LP.
http://www.epd.com/power meters.html#PLM-1.
Several avenues exist to improve Gemini. First, an ad- [6] PowerCost Monitor.
justable burden resistor would allow for a wider range of loads http://www.bluelineinnovations.com/powercost-monitor-2.
[7] Smarthome iMeter Solo. http://www.smarthome.com/
to be metered. Determining the correct value would entail a 2423A1/iMeter-Solo-INSTEON-Power-Meter-Plug-In/p.aspx.
dynamic search to scale the current transformer output voltage [8] TED: The Energy Detective.
for the ADC reference, or vice versa. http://www.theenergydetective.com/.
[9] Texas Instruments CC2538.
Second, Gemini could improve its voltage waveform synthe- http://www.ti.com/product/cc2538.
sis for calculating power. The voltage monitor could compute [10] Watts Up? .Net. https://www.wattsupmeters.com/secure/
and transmit additional Fourier coefficients and the power products.php?pn=0&wai=0&spec=1.
[11] Department of Energy (DOE) Buildings Energy Data Book.
meter could dynamically create the voltage waveform instead http://buildingsdatabook.eren.doe.gov/, 2012.
of using a precomputed lookup table. [12] Analog Devices. ADE7753 Datasheet. http://www.analog.
Third, Gemini could be extended to calculate energy as com/static/imported-files/data sheets/ADE7753.pdf.
[13] S. DeBruin, B. Campbell, and P. Dutta. Monjolo: An
well as power. To calculate energy, a power meter needs to energy-harvesting energy meter architecture. In Proceedings of
time the interval between power measurements, which it could the 11th ACM Conference on Embedded Networked Sensor
do by requesting time information from the voltage monitor Systems, SenSys ’13, 2013.
[14] Department of Energy. Building energy efficiency frontiers and
when it requests the voltage parameters. Then a power meter incubator technologies (BENEFIT), 2014.
could multiply the elapsed time by the power measurements to [15] P. Dutta, J. Taneja, J. Jeong, X. Jiang, and D. Culler. A
calculate energy. This would be particularly useful for using building block approach to sensornet systems. In SenSys’08: In
Proceedings of the Sixth ACM Conference on Embedded
Gemini as a whole-house meter to corroborate revenue-grade, Networked Sensor Systems, Nov. 2008.
utility-installed meters. [16] X. Jiang, S. Dawson-Haggerty, P. Dutta, and D. Culler.
Design and implementation of a high-fidelity AC metering
VIII. C ONCLUSION network. In IPSN ’09: Proceedings of the 2009 International
Existing power meters are limited in their ability to solve Conference on Information Processing in Sensor Networks,
pages 253–264, Apr. 2009.
the critical problem of electrical submetering in buildings. [17] J. Lifton, M. Feldmeier, Y. Ono, C. Lewis, and J. A. Paradiso.
Installation challenges, maintenance headaches, and metering A platform for ubiquitous sensor deployment in occupational
accuracy hamper the widespread adoption of submetering. and domestic environments. In IPSN ’07: Proceedings of the
6th international conference on Information processing in
To address these challenges, we present Gemini, a new true sensor networks, Cambridge, Massachusetts, Apr. 2007.
power meter that offers clip-on installation, energy-harvesting [18] Linear Technology. LTC3588 datasheet.
operation, and accurate power measurement. Our work draws http://cds.linear.com/docs/en/datasheet/35881fa.pdf.
[19] National Science and Technology Council–Committee on
on two complementary research themes—virtual sensors and Technology. Submetering of building energy and water usage:
energy-harvesting meters—bringing them together to yield a Analysis and recommendations of the subcommitte on
novel design point in the metering space. Gemini addresses the buildings technology research and development, Oct. 2011.
[20] S. N. Patel, S. Gupta, and M. S. Reynolds. The design and
drawbacks with prior approaches by decoupling and distribut- evaluation of an end-user-deployable, whole house, contactless
ing the AC voltage and current measurement acquisitions, and power consumption sensor. In Proc. of the SIGCHI Conference
recombining them over a low-bandwidth, time-synchronized, on Human Factors in Computing Systems, 2010.
[21] N. B. Priyantha, A. Kansal, M. Goraczko, and F. Zhao. Tiny
wireless channel, to offer non-invasive real, reactive, and web services: design and implementation of interoperable and
apparent power metering. This represents a significant ad- evolvable sensor networks. In Proceedings of the 6th ACM
vancement from previous designs that are difficult to install, conference on Embedded network sensor systems, 2008.
[22] T. Schmid, D. Culler, and P. Dutta. Meter any wire, anywhere
require periodic battery maintenance, or cannot calculate true by virtualizing the voltage channel. In Proceedings of the 2nd
power. By offering comparable accuracy to commercial meters ACM Workshop on Embedded Sensing Systems for
while vastly improving deployability, Gemini offers a new tool Energy-Efficiency in Building, pages 25–30. ACM, 2010.
[23] U.S. NSB. Building a sustainable energy future: U.S. actions
with which to attack the building submetering problem. for an effective energy economy transformation, Aug. 2009.
[24] T. Weng, B. Balaji, S. Dutta, R. Gupta, and Y. Agarwal.
IX. ACKNOWLEDGMENTS Managing plug-loads for demand response within buildings. In
This work was supported in part by the TerraSwarm Research Proceedings of the Third ACM Workshop on Embedded
Center, one of six centers supported by the STARnet phase of the Sensing Systems for Energy-Efficiency in Buildings, pages
Focus Center Research Program (FCRP), a Semiconductor Research 13–18. ACM, 2011.
Corporation program sponsored by MARCO and DARPA. This [25] Q. Xu, I. Paprotny, M. Seidel, R. White, and P. Wright.
material is based upon work partially supported by the National Stick-on piezoelectromagnetic AC current monitoring of circuit
Science Foundation under grants CNS-0964120, CNS-1111541, and breaker panels. Sensors, 13(3), 2013.
CNS-1350967, and generous gifts from Intel and Texas Instruments.