IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
Environment and Application Testbed for
Low-Power Energy Harvesting System Design
Lukas Sigrist, Andres Gomez, Matthias Leubin, Jan Beutel, Member, IEEE and Lothar Thiele, Member, IEEE
Abstract—Energy harvesting systems strongly depend
on the non-deterministic behavior of the environment. Sys-
tematic and thorough evaluation of these systems demands
for tools that consistently reproduce these conditions and
allow for closely integrating highly dynamic applications.
To this end, a testbed is introduced that allows to precisely
and repeatedly force an energy harvesting system under
test using thermal and radiative sources in a controlled
environment while allowing to sinking arbitrary current pro-
files. The coordinated control of the boundary conditions
on the in- and output side enables detailed evaluation, ex-
ploration and dimensioning of different aspects of energy
management and harvesting systems. By reproducing en-
vironmental traces at a higher rate than normally occurring
in nature, the testbed allows to substantially shorten the
time needed for experimental evaluations. This approach
enables fast and consistent evaluation of energy harvesting
systems under a wide coverage of operating conditions.
Index Terms—Emulation, Energy Harvesting, Environ-
ment, System Design, Testbed
I. I NTRODUCTION
Testbeds are an invaluable tool for evaluating systems under
controlled conditions. In particular in contexts where the
primary factors affecting the system performance exhibit sig-
nificant non-determinism or variability, they enable controlled
system evaluation and detailed analysis of the impact these
factors have. In the domain of low-power energy harvesting
systems, jointly managing the energy extraction from a non-
deterministic environment and satisfying the dynamic load
requirements represents a major challenge. Consequently, en-
ergy management solutions addressing these challenges need
to adapt to highly dynamic operating conditions to operate
efficiently. This has also implications on their evaluation: for
thorough characterization, analysis, dimensioning and valida-
tion of these systems, energy harvesting systems need to be
evaluated with full coverage of the multi-dimensional and
widely varying operating conditions. In addition to these
requirements in terms of dynamic range and precision, the
interactions of the testbed with the target system must be time-
synchronized. Therefore, an appropriate coordination though a
global control scheme is required. For a feedback-based design
Manuscript received May 19, 2020; revised September 16, 2020;
accepted October 18, 2020.
This research was funded by the Swiss National Science Foundation
under grant 157048: Transient Computing Systems.
Lukas Sigrist, Andres Gomez, Matthias Leubin, and Lothar Thiele are
with the Computer Engineering Group, ETH Zurich, Zurich, Switzerland.
Jan Beutel is with the Computer Engineering Group, ETH Zurich, Zurich,
Switzerland and the Department of Computer Science, University of
Innsbruck, 6020 Innsbruck, Austria.
flow to benefit from these evaluations, the conditions under
which they are performed have to be consistent to guarantee
comparability among them. Furthermore, rapid evaluation is
required to provide timely feedback in the design flow. Con-
sequently, the interactions between the testbed and the system
under test need to be accelerated, in terms of generation of
the physical environment, consumption of energy, as well as
in terms of measurements.
In summary, a testbed for energy harvesting systems has to
address the issues related to the environment’s reproducibility
and wide-ranging electrical load properties, while providing
accelerated evaluation of multi-dimensional configurations. To
the best of the authors’ knowledge, there are no integrated
testbeds for low-power energy harvesting systems available
that satisfy the above requirements in terms of flexibility, mod-
ularity with respect to harvesting technology, high dynamic
range and precision. Building such a testing environment from
commercial off-the-shelf equipment does not only come with
high costs but also hampers reproducability of results due to
its ad-hoc composition and need for extensive calibration. In
addition, actuation and measurement components need to be
tightly controlled in order to guarantee time-synchronization
and to allow for accelerated evaluation using time-scaling.
This work addresses these challenges in evaluating energy
harvesting systems by introducing an integrated testbed that
provides controlled emulation of the environment’s physical
properties and the application requirements with a high dy-
namic range, combined with high accuracy in terms of actu-
ation, measurements, and time synchronization. Specifically,
the contributions are summarized as follows:
• The first integrated testbed for low-power energy har-
vesting systems that combines high-dynamic range envi-
ronment and load actuation with comprehensive system
observability is designed, implemented and characterized.
• Two environment emulators that provide consistent re-
production of thermal and visible light properties of the
environment are introduced. This enables in-depth analy-
sis of thermoelectric and photovoltaic energy harvesting
systems with full coverage of their harvesting conditions.
• A centralized feedback controller allows for automated
measurement campaigns, and enables the speed-up of ex-
haustive evaluations of energy harvesting systems through
time scaling and a high degree of time-synchronization.
The presented testbed solution provides fast and consistent
evaluation with unprecedented coverage. Three case stud-
ies highlight the significantly simplified, automated and fast
experimental evaluation in the design of low-power energy
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
harvesting systems.
The remainder of this paper is organized as follows. Related
work is covered in Section II. After introducing the overall
testbed concept in Section III, the environment emulation,
the load emulation and the control system are described in
Sections IV, V and VI, respectively. The benefits of the
testbed implementation are highlighted in two case studies in
Section VII, before concluding in Section VIII.
II. R ELATED W ORK
In this section, previous work related to the emulation of
the environment’s physical properties and energy harvesting
sources in general are covered. Further, solutions presented in
the area of reproducing electrical loads with dynamic behavior
are reviewed.
Environment Emulation: For photovoltaic and thermo-
electric harvesting considered in this work, the focus lies
on the emulation of temperature gradients and solar radia-
tion. An approach to characterizing thermoelectric generators
(TEGs) taken in [1] was to experimentally determine their
electrical and thermal equivalent model parameters, and to
investigate their relationship. Alternatively, the I-V-character-
istics of TEGs were extracted by varying their temperature
gradient and connecting different power resistors as load
[2]. Wang et al. recently proposed a circuit for combined
harvesting from thermoelectric and piezoelectric transducers
[3]. In their evaluation, they relied on a testbed consisting
of a shaking table for the piezoelectric, as well as hot/cold
containers and hot-dissipating metal sheets for thermoelectric
harvesting. In the characterization of photovoltaic materials,
accurate reproduction of the solar spectrum is key to perform
low-level characterization of photovoltaic cells like measuring
quantum efficiency [4]. To this end, solar simulators have
been developed for large scale concentrator photovoltaic sys-
tems [5]. Similarly, highly accurate commercial sun spectrum
sources exist for device testing [6], but are very costly with
over 80 kUSD for an illumination area of 45 cm × 45 cm.
Low-cost alternatives using arrays of colored LEDs and a
well characterized spectrum have been proposed by Bazzi
et al. for testing photovoltaic cells [7]. Recent work has
also studied the effects of different spectra and intensities
in indoor photovoltaics, which can experience a mixture of
solar and artificial light sources [8]. Similar to previous
proposals, this work relies on dual-sided temperature forcing
to reproduce consistent temperature gradients for evaluating
thermoelectric harvesting. Unlike the characterization testbeds
for photovoltaic cell materials, the novel testbed emulates
diverse photovoltaic harvesting scenarios with illuminance
levels ranging from dimly lit indoor environments to direct
sunlight.
Energy Harvesting Emulation: In many harvesting-based
works, spectrum dependencies are often abstracted away, as
the electrical energy input to the system is typically the
decisive factor. In the high power domain, high-end solar
array simulators (SASs) can emulate user-programmable I-
V curves for different irradiance and temperature conditions
[9]. These SASs are then used to facilitate the development
of novel maximum power point tracking (MPPT) algorithms
under reproducible conditions [10]. In the low power domain,
emphasis is put on either reproducing illuminance conditions
[11], [12] or electrical circuits exhibiting similar behavior [13].
Other works abstracted away the physical environment and
transducers and directly emulate their electrical behavior [14]–
[16], or control the input of harvesting circuits to force them
to recorded operating points [17]. In this work the focus is on
emulating the physical properties of the environment as they
present the original source of the non-determinism introduced
into energy harvesting systems.
Electrical Load Emulation: Electronic current sinks are
available in various forms. Bench-top source measure units
(SMUs) [18] can sink currents with high accuracy and are
programmable. There also several proposals of custom-built,
configurable loads. Jian et al. designed an interleaved multi-
phase pulse width modulation (PWM) scheme to mitigate rip-
ples in power consumption [19]. In [20] the author presented
a programmable DC load with different operation modes to
emulate arbitrary load profiles. Upadhayay et al. developed a
switching-converter-based electronic load with a high band-
width [21]. Programmable DC loads can be used in battery-
lifetime estimation [22], and transducer characterization [8]. In
RF harvesting circuits, low power resistor emulation can also
be used for MPPT [23]. While most of the aforementioned
works report high precision, they aimed at either high power
or low power electronic loads. With focus on the embedded
energy harvesting domain, the emulator introduced in this
work aims to represent embedded system profiles, which
combine elements of both. It can reproduce currents ranging
from a few nA to peak active currents in the 100’s mA range.
III. R EQUIREMENTS AND S YSTEM OVERVIEW
This section starts with summarizing the requirements for
the physical environment emulation. Afterwards, an overview
about the proposed testbed and its components is given.
A. Testbed Requirements
The characterization, design and testing of low-power sys-
tems that draw all or some of their energy from the envi-
ronment is challenging due to the close coupling of system
components like harvesting devices, energy management, pro-
cessing hardware and software. In addition, the high dynamic
range of electrical quantities due to changing environmental
conditions as well as non-deterministic load requirements
poses additional challenges for analysis, dimensioning and
validation of these systems. Therefore, the following major
requirements for a suitable testing strategy and the correspond-
ing measurement support can be identified:
(i) The characterization and test of individual system com-
ponents as well as their interaction in the overall system
must be supported.
(ii) The energy harvesting as well as the electric load can
be evaluated with full coverage of the multi-dimensional
and widely varying operating conditions.
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
System under Test RocketLogger
Environment
Emulation SmartLoad
Environment
Traces Centralized
Emulation
Power Controller Measurement
Traces Traces
Figure 1. Overview of the integrated testbed including emulation, mea-
surement, input and output traces as well as the centralized controller.
(iii) The measurements should be sufficiently accurate and
reproducible, but at the same time allow for testing
realistic deployment scenarios in a reasonably short time.
The overall design of the testbed architecture and the
associated testing strategy satisfies these requirements and
provides viable answers to the associated challenges.
B. System Overview
Figure 1 depicts the main components of the proposed
integrated testbed for low-power energy harvesting systems. Its
individual components and controllers, including environment
and load emulation, will be described in detail in the forth-
coming sections. Following requirement (i), the system under
test can be exposed to realistic harvesting environments for
visible light or thermal gradients, see Section IV. In parallel,
a programmable smart current sink allows the emulation of
the energy requirements of an application that could run on
the system under test, see Section V. The stateful, time-
and event-triggered emulation of harvesting as well as load
traces allow for a realistic characterization and validation of
all individual system components as well as their interaction.
The RocketLogger component [24] is a mixed-signal tracing
platform responsible for measuring electrical quantities of
the system under test and the testbed. Its measurements are
used for detailed observation of the system under test and
to provide input to the centralized emulation controller. The
RocketLogger, the environment emulation and the smart load
are integrated seamlessly to support a wide range of multi-
dimensional operation conditions, as demanded by require-
ment (ii). The emulation controller supports automated and
exhaustive evaluation of multi-dimensional configurations, see
Section VI. It coordinates the operation of all components
of the testbed and allows to explore the system performance
at predefined setpoints and to replay stored traces for the
harvesting environment as well as the electric load. Due to
the embedded control algorithms, an accurate synchronization
between the availability of the emulated energy, the energy
use of the load and the associated measurements is provided.
Satisfying requirement (iii), the high time precision also allows
time scaling to accelerate the system evaluation experiments.
IV. P HYSICAL E NVIRONMENT E MULATION
The emulation of the environment’s physical properties
needs to offer a high accuracy to provide consistency across
Setpoint Control
and Measurement
Controller Controller
USB USB
force sense Distance Adjustment sense force
Peltier TEG/Thermal System Peltier
Element Under Test Element
Figure 2. The architecture of the thermal environment evaluation with
two software controlled Peltier elements for consistent temperature
gradient emulation. The adjustable distance enables evaluation of stan-
dalone thermoelectric generators (TEGs) or complete thermal systems
including thermal guides.
Figure 3. The thermal environment emulation. A thermal harvesting
system consisting of a thermoelectric generator (TEG) and an attached
copper thermal guide is evaluated in between the two Peltier elements.
experimental evaluations. This enables a comparison of the
system performance under the same environmental conditions.
In addition, the high dynamic range and time variability of
the harvesting environment provides major challenges for the
emulation of the relevant physical properties, such as thermal
gradients and visible light.
A. Thermal Gradient Emulation
The thermal environment emulation component of the
testbed consists of two digitally controlled Peltier elements
that offer cooling and heating of a metal plate terminal. Fig-
ure 2 illustrates the architecture of this emulation component.
Commercial CP-061HT Peltier elements from TE Technology
are employed for this purpose and the corresponding TC-
720 unit is used for control. This unit integrates feedback
control of the Peltier element’s temperature and measures
the temperature of the forced terminal. Setpoint configura-
tion and measurement readout are exposed through a USB
interface that allows the logical integration into the testbed
architecture according to Figure 1. As shown in Figure 3,
these two Peltier elements are installed on a custom built
frame with the temperature-controllable terminals facing each
other. The adjustable distance between them allows inserting
thermal harvesters of various sizes. This enables evaluation of
standalone TEGs, as well as their integration into a thermal
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
LED LED Dimmer
Mirror Walls
Photovoltaic Cell
Current Extraction
sense
Illuminance Controller
Sensor Peltier force
Element
Figure 4. The proposed solar emulator consists of a dimmable LED light
source, a mirror box for a uniform illuminance distribution, and a Peltier
element to control the photovoltaic cell’s temperature. Temperature and
illuminance sensors enable logging of the emulated conditions.
system including thermal guides. The energy transfer is limited
by that of the Peltier elements, which can extract/insert up
to 63 W at a 0 °C temperature difference to the ambient. In
combination with the external controller, this enables accurate
and fast transitions between temperature setpoint.
B. Visible Light Emulation
In contrast to the thermal emulation with its logical in-
tegration into the testbed, see Figure 1, the solar environ-
ment emulator is also physically integrated into the testbed
as illustrated in Figure 4. This custom-built solar emulator
enables consistent emulation of illuminance conditions ranging
from dimly lit indoor environments to direct sunlight. In the
following, the most important design aspects of the emulator
as integrated into the testbed shown in Figure 5 are discussed.
Light Source: The light source in the testbed consists of
light-emitting diodes (LEDs). They are compact, cost effective
and their illuminance is well controllable over a large range. In
addition, their short warm-up time guarantees a fast and accu-
rate response during emulation. The high efficiency of LEDs
also simplifies heat management when emulating the high
illuminance levels of direct sunlight. On the downside, LEDs
cannot accurately reproduce the spectrum of the sunlight.
However, as was show in [7], this approach could be extended
to an array of colored LEDs to improve the reproduction of the
spectrum. In the emulator, Aventrix 2x2 matrices consisting of
four NVSWE21A LEDs were employed as light source. Their
high efficiency and luminous flux enable emulation of direct
sun light conditions. A total of nine modules were mounted
on a heat sink to cover a test area of 180 mm × 180 mm. To
avoid high supply voltages, the control of these LED modules
is distributed to four dimmer circuits that are discussed below.
Illuminance Uniformity: For consistent evaluation of
photovoltaic cells it is important to illuminate them uniformly
[25]. Simulation of the illuminance distribution showed that
a uniform illuminance distribution is achieved when the LED
light sources are arranged in an evenly spaced grid and the
light is reflected at the boundary of the light box. Conse-
quently, mirrors are used for shielding the light box. Illumi-
nance measurements with only one of the mirrors removed
confirmed that the illuminance differs up to 10 % across the
Figure 5. The seamlessly integrated components of the custom-
developed solar testbed. The top half represents the mirror-shielded
light box with the LEDs mounted below the heat sink. The bottom of
the light box consists of a Peltier controlled aluminum plate to maintain
a constant temperature of the photovoltaic cell during the evaluations
(hidden in the shadows). On the right side of the testbed the control
circuitry is mounted: the LED dimmer with high dynamic range on
top, the SmartLoad (presented in Section V) on the bottom left, and
the RocketLogger platform providing the measurement capabilities and
hosting the testbed control (discussed in Section VI) on the bottom right.
R1
Setpoint
Configuration Step-Down Current
I2 C Current Regulation
Range Regulator
M1
LED (ext.)
Controller PWM Dimming D1
Cout
M2
Output Filter
Figure 6. To provide control over a large illuminance range from 10 lx
to 120 000 lx, a custom two-stage LED dimmer circuit was developed.
It combines step-down current regulation and pulse width modulation
(PWM) dimming. Multiple of these channels are provided for controlling
a total of 36 power LEDs.
light box, while the configuration with dual sided mirrors
exhibited only minimal variability.
LED Illuminance Control: The illuminance of the LEDs
is controlled by either adjusting their supply current, or using
a constant current source and pulse width modulation (PWM).
There exist control modules for either approach, but they
are typically not made for high-accuracy dimming and their
precision is often not specified. Therefore, a custom LED
driver was implemented to provide accurate control of a wide
illuminance range. The dimmer circuit shown in Figure 6
combines a constant step-down current regulation using an
LM3409HV buck controller and PWM dimming. To provide
a high dynamic range, the step-down regulator is operated
at two different currents, and PWM dimming is used for
detailed illuminance control. Unfortunately, PWM dimming
introduces ripples in the output current that could be captured
by photovoltaic cells. To reduce this undesired effect to a
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
4 DAC
Relative error [%]
Emulation error
3 State Machine
2 Current Trace
1 I2 C
0 AV = 100
101 102 103 104 105
Illiminance level [lx]
Figure 7. The illuminance emulation shows a high accuracy of 8.8 lx +
0.5 % over the range from 10 lx to 120 klx. A significant increase in
emulation error below 100 lx shows the limitations of the presented
emulator for extremely low illuminance levels.
minimum, a high modulation frequency and output filtering is
used. The step down regulator’s range and the PWM signal are
controlled by a microcontroller that exposes an I2 C interface
for digital control of the illuminance level.
Temperature Control: As the performance of photovoltaic
cells is considerably affected by the ambient temperature
[26], a CP-036HT Peltier element is included to control the
temperature of the panel inside the light box. The same TC-
720 controller used in the thermal environmental emulator is
employed for precise temperature control and measurement.
This guarantees consistent evaluation of solar harvesting when
the LEDs emit significant amounts of heat at high illuminance
levels, and enables temperature dependent performance anal-
ysis of photovoltaic cells.
Environmental Condition Measurement: The emulated
illuminance level and the temperature are monitored for the
purpose of dimmer calibration and feedback on the emulation
consistency when traces with fast transients are emulated
(see also the discussion regarding accelerated emulation in
Section VI-B). The temperature of the emulator plate is read
out from the Peltier controller that measures this temperature
for feedback control. Four TSL4531 ambient light sensors are
placed at the edges of the illuminated test area to measure
the illuminance. With a wide measurement range from 3 lx
to 220 klx and a specified sensitivity of 1 lx, these sensors
accurately cover the illuminance range emulated in the testbed.
C. Characterization of the Solar Environmental Emulator
The accuracy of the solar emulator is evaluated to confirm
that it satisfies the requirements of consistent emulation of a
wide range of illuminance levels. Prior to characterization, it
was calibrated in a two-step approach using the integrated light
sensors as a reference. First, the LED output illuminance EV
is translated into an LED current ILED using a linear model.
Subsequently, the duty cycle for each current channel/range
combination is calibrated using a 3rd order model to match
the current ILED . The high efficiency of the LEDs enables the
emulation of illuminances from 10 lx up to 230 klx, which
represents an illuminance higher than that direct sunlight
(120 klx). With this coverage, photovoltaic harvesting sce-
narios with very high dynamic range can be emulated. The
accuracy of the emulation is evaluated by sweeping the full
illuminance range. Comparing the setpoints of the sweep to
Controller
M1
AV Rlo Rmed Rhi V
Triggers Rsh
Control
Figure 8. High-level schematic of the SmartLoad programmable current
sink. A feedback controlled MOSFET sinks a current that is set by
the digital-to-analog converter (DAC). The controller exposes an I2 C
interface for configuration and replays current traces in time- and event-
triggered manner.
the sensor measurements results in the relative emulation
errors shown in Figure 7. These results indeed confirm a
high accuracy of the emulated illuminance of 8.8 lx + 0.5 %.
A noticeable increase in error is however reported for very
low illuminance levels below 100 lx. This is predominantly
attributed to the low PWM frequency that had to be chosen
for the low illuminance range to prevent the step-down current
regulator from overheating. In a redesign with stricter focus
on thermal design, this can likely be resolved and the accuracy
improved.
V. E LECTRICAL L OAD E MULATION
This section introduces the SmartLoad component of the
integrated testbed, see Figure 1, a programmable current
sink that emulates the application’s energy requirements. The
stateful, time- and event-triggered emulation of current traces
enables replaying of synthetic or recorded current consumption
behavior of low-power systems.
Similar to the environment emulation, the SmartLoad has to
provide consistent emulation, while covering the full range of
current consumption in today’s low-power systems that range
from nA to 100’s mA. The emulation requires update rates
and response times of typical embedded system in order to
faithfully reproduce energy consumption traces. To employ the
SmartLoad for transducer characterization, it has to support
accurate current sinking at zero input voltage to allow for short
circuit current characterizations.
A. SmartLoad Architecture
The architecture of the SmartLoad is illustrated in Figure 8.
In the following, the analog current sink circuit and the digital
control used for accurate replay of current traces are covered
in detail.
Current Sink: The analog front-end uses a
CSD18542KTT power MOSFET M1 as controlled current
sink. The feedback measurement from the shunt resistor Rsh
is amplified by an AD8421 instrumentation amplifier. This
signal is compared to the reference level set by an AD5761R
digital-to-analog converter (DAC). The amplified signal
difference provided by an OPA130 operational amplifier
(op-amp) controls the MOSFET M1 . For precise current
control and to provide a high dynamic range, the shunt
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Table I Table II
DC P ERFORMANCE OF THE S MART L OAD DYNAMIC P ERFORMANCE OF THE S MART L OAD.

Name Range Accuracy Range % of Step Size Time

High Range 5 mA to 500 mA 0.01 % + 80 µA 90 % 32 µs
High Range
Medium Range 50 µA to 5 mA 0.008 % + 100 nA 99 % 173 µs
Low Range 0 µA to 50 µA 0.003 % + 2.25 nA
Medium 90 % 15 µs
Range 99 % 138 µs

resistor value Rsh used for feedback measurement can be 90 % 145 µs

Low Range
switched between three values. This enables more precise 99 % 340 µs
control using the DAC and at the same time keeps the burden
voltage at the input terminals low. An offset voltage of 1 V 10.6 5
can be inserted into the current path to enable current sinking SmartLoad current
at zero input voltage. This is important for current sinking 10.4 Input voltage 4

Current [µA]

from low voltage sources such as thermoelectric generators
(TEGs) or to perform short circuit current characterization of
transducers.
Control Loop Stability: The feedback control circuit
was modeled to optimize it for stable operation and highly
accurate current control over the full operating range. Using
the device-specific transfer functions of the op-amp, MOSFET,
and amplified shunt resistor measurement, the circuit’s stability
and output offset error were analyzed. A high feedback gain of
AV = 100 for the linear amplifier was selected to reduce the
offset error in the output current. To increase the very small
stability margin at this amplification, a phase lag compensator
was added in front of the power MOSFET to obtain a robust
current control (not shown in Figure 8).
Event-Triggered Emulation: The digital-to-analog con-
verter (DAC) is controlled by a microcontroller to provide
stateful replay of stored current traces. Using the I2 C interface,
current traces are downloaded to the SmartLoad, and the replay
of them is configured and controlled. In the present implemen-
tation, a replay is triggered through the control interface or
external triggers. However, the functionality of the SmartLoad
controller is extendable to support more complex state ma-
chines and state-dependent digital outputs. Such extensions
allow for digital reconfiguration of the energy management
dependent on the application state.
B. Characterization of the Electrical Load
While the requirement of sinking short circuit currents
at zero input voltage is guaranteed by design, the accuracy
and dynamic response behavior of the SmartLoad was eval-
uated experimentally. Prior to evaluating its performance, the
SmartLoad was calibrated. A Keithley 2450 source measure
unit (SMU) was used to source the current and to preform
reference measurements. Current sweeps were performed for
all three current ranges to determine their individual linear
calibration parameters.
DC Accuracy: Table I shows the DC performance of
the SmartLoad 4 h after calibration. This evaluation was
performed again using the SMU as source and for reference
measurements at an input voltage of 0 V. The results show a
very high accuracy of the sunk current for all three ranges.
Higher input voltages likely result in a larger offset error in
Voltage [V]
3
10.2
2
10
1
9.8 0
0 1 2 3 4 5 6
Time [s]
Figure 9. The input voltage step rejection evaluation of the SmartLoad
shows short current spikes of about 0.3 µA at the lowest current range.
the lowest range, since the leakage current of the MOSFET
will be in the range of several 10 nA.
Dynamic Behavior: Table II shows the worst case step
response times of the SmartLoad when performing a full range
current step. The values shown represent the delay from the
time a new setpoint was set until the output has reached 90 %
or 99 % of the step size. In summary, the maximum delay
of the SmartLoad is less than 400 µs if switching the current
range. This performance is sufficient to faithfully emulate the
behavior of typical low-power systems.
Input Voltage Step Rejection: The last parameter evalu-
ated is the SmartLoad’s behavior under input voltage steps
from 0 V to 5 V forced by the SMU. The result of this
experiment when sinking a current of 10 µA at the lowest
range is shown in Figure 9. Only very short spikes of about
0.3 µA for less than 2 ms were observed in this current range,
as the current sink’s feedback loop has to adapt to the changed
operating point. In the high current range these spikes were
so small that they disappeared in the measurement noise of
the SMU. This input voltage rejection is strong enough to
guarantee repeatable and accurate emulation of the current
consumption under the voltage variability of typical energy
harvesting systems. In specialized use cases with highly dy-
namic voltage supplies, the feedback loop can be optimized
further.
VI. C ENTRALIZED E MULATION C ONTROL
To satisfy the requirement of seamless integration and
joint control of the physical emulators and the SmartLoad,
these components are integrated in an overall testbed. Global
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
control enables automated and exhaustive evaluation of multi-
dimensional configurations. Scaling of the time domain accel-
erates the evaluation to provide accelerated system evaluation,
but demands careful consideration of potential side effects.
A. Automated Emulation Control
The emulation of the physical environment and of elec-
trical consumers is combined with the mixed-signal tracing
capabilities of the RocketLogger platform [24]. The extensible
architecture of the logger is leveraged to implement the mech-
anisms for central control of the emulation and measurements.
Time series inputs specify which environment parameters and
current consumption to emulate, as is illustrated in Figure 1.
During the experiments, various energy flows, environment
conditions and system states are traced. This enables detailed
analysis of the system performance and behavior of the energy
harvesting system.
B. Controlling the Time Domain
To accelerate the experimental evaluation, the time domain
of the emulated traces can be scaled. Time scaling with a user-
defined procedure is performed directly on the timed traces of
the system. This flexibility allows to not only scale with a
fixed speed-up factor, but also selectively using time-variant
scaling methods. The accelerated replay of these traces leads
to much faster transients that need to be emulated. Therefore, it
is crucial to consider that these transients can only be emulated
within given bounds (see the step response characterizations
in Sections IV and V). Developing a general methodology to
determine the speed-up factor is difficult for many reasons: the
parameters of the system under test like thermal and electrical
capacities, or fast transients in the recorded trace can result
in faster dynamics than the testbed can emulate. This can
significantly distort the experimental evaluation and may even
lead to invalid conclusions. However, tracing the emulated
conditions with the integrated sensors enables their compari-
son to the specified input traces and allows to validate whether
the dynamics of the trace were emulated successfully. While
incorporating the sensor feedback ensures valid environment
emulation, a similar mechanism is necessary to validate the
correct behavior of the energy harvesting system under test.
C. Rapid Setpoint Characterization
A typical task in system design is evaluating a large number
of setpoints for design space exploration. Therefore, rapid
setpoint characterization is important for high coverage and
fast feedback. However, changing too rapidly between the set-
points can seriously distort the measurements due to transient
effects. An analysis highlighting this issue for an AM-5412
photovoltaic cell characterization is shown in Figure 10: it
shows the deviation of the measurements from the long-term
average within a 1000 ms window starting at the specified
delay after the setpoint update. This can lead to a significantly
increased error if performing the analysis too early. Similar
behavior has been observed when performing thermoelectric
generator (TEG) characterization and switching the test point
2.5
Relative error
2
Error [%]
1.5
1
0.5
0
100 101 102 103 104
Delay after setpoint update [ms]
Figure 10. The setpoint characterization measurement of an AM-5412
photovoltaic cell shows drastically increased error if the analysis is
performed using data points too close to the setpoint change.
too quickly. Due to the much slower dynamics of thermal sys-
tems, the average settling time when adjusting TEG setpoints
is several minutes. To prevent these errors from happening, the
testbed control provides a setpoint characterization mode. In
this mode, the testbed control performs steady-state detection
based on the algorithm presented in [27]. Using this feedback
mechanism, the characterization is automatically accelerated
in a controlled manner. A system designer only provides the
list of setpoints instead of time series data and the signal and
analysis time window to use for steady-state detection. The
time window is dependent on the system dynamics, but could
be identified using a setpoint update experiment similar to the
analysis shown in Figure 10. This highly-effective procedure
is demonstrated in the transducer characterization case-study
in Section VII-A.
VII. C ASE S TUDIES
Three case studies demonstrate how the testbed supports
rapid evaluation in various aspects of the energy harvesting
system design. First, the automatic characterization thermo-
electric generators (TEGs) and photovoltaic cells is performed.
The second case study highlights that fast and accurate evalua-
tions are possible by speeding up the evaluation. The final case
study focuses on the evaluation of an energy-driven execution
using an event-triggered application.
A. Automated Transducer Characterization
Thanks to the fully-integrated harvesting testbed and its
centralized control, the characterization of the transducers
requires very little effort: exposing the transducer to the
desired operation conditions in the thermal or solar testbed and
connecting its output to the SmartLoad are all the necessary
steps. The testbed control automatically handles sweep and
steady state detection of the specified list of setpoints. Fig-
ure 11 shows an exhaustive multi-dimensional characterization
of an AM-5412 photovoltaic cell, i.e., the transducers output
power as a function of the illuminance and voltage levels of
the emulated operating point. The cell was characterized for
the full range from an illuminance level of 100 lx (dimly lit
indoor) to 120 klx (direct sunlight), and from open-circuit to
short-circuit conditions. The plot consists of a total of 2160
characterization points that were recorded in less than 80 min.
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

50 Table III
E VALUATION S PEED -U P AND ACCURACY.
40 S PEED - UP FACTORS HIGHER THAN 20× ARE CONSIDERED UNSAFE .

Power [mW]
30 Speed-Up Duration Nmeasured Nexpect Error
20 1× 46 h 33091 baseline –
10 2× 23 h 16459 16546 −0.52 %
5× 9h 6492 6618 −1.91 %
0 10× 4.5 min 3229 3309 −2.42 %
20× 137 min 1627 1655 −1.67 %
125
100 50× 55 min 603 662 −8.89 %
75
0 0.5
50 100× 27 min 183 331 −44.70 %
1 1.5 2 25
2.5 3 3.5 0
Voltage [mV] Illuminance [klx]
optimize the energy management system.
Figure 11. The automated transducer characterization enables rapid
and exhaustive evaluation over a wide range of operating points. The B. Rapid System Evaluation
illustrated characterization of an AM-5412 photovoltaic cell at a tem-
perature of 25 °C consists of 2160 setpoints that were recorded in only This case study demonstrates the feasibility of accelerated
76.6 min. system evaluation by emulating real-world illuminance traces
at much higher rate than they occur in nature. It is con-
7 firmed that accurate results are obtained when increasing the
6 trace replay rate, thus enabling significantly shorter evaluation
5 times. In general, the limits of the speed up may depend on
Power [mW]

all components of the system under evaluation, for example

4
the physical harvesting mechanism, the sensor, the energy
3
management and the application that is running on the energy
2 consumer. In case there is no information about the imposed
1 speed limits, then an iterative approach is a safe way to
0 determine bounds on the achievable acceleration factor.
10 The experiment uses a 46 h long illuminance trace that was
8 recorded close to a window in an office scenario, where a
6
4 mixture of artificial and indirect natural light can be exploited
0 100 200 300 2
400 500 600 0 for harvesting. A Bluetooth Low Energy (BLE) sensor tag
700 Gradient [K] consisting of an AM-5412 solar panel and an AEM10941
Voltage [mV]
harvesting chip that supply a TI CC2650 system on a chip
Figure 12. The thermal capacitances of the testbed and limited output
power of the testbed demand for 30 h for a 461 setpoint characterization (SoC) is used as sample application. For this case study,
of a thermoelectric generator (TEG). The steady state detection guaran- the performance of the BLE tag, characterized by the total
tees accurate measurements without transient effects originating from number of executions N , is analyzed. The baseline trace is
setpoint switching.
replayed at rates that match that of the recording (1×), and
up to 100 times faster (100× speed-up). The experimental
Similarly, Figure 12 shows the characterization performed for results are summarized in Table III. The number of executions
a TEG241-150-29 TEG. Due to the thermal mass of the entire Nmeasured for the accelerated trace replay (2 up to 100×) are
system and the limited output power of the Peltier element, compared to the baseline (1×) where the trace is replayed
the delay for reaching a stable setpoint measurement is much at the recording rate. The accuracy of the experiment is
higher. For this reason the recording of the 461 setpoints evaluated by comparing Nmeasured to the theoretically-scaled
took about 30 h. However, the fully automated process allows baseline value Nexpected . At extreme speed-up factors beyond
the characterization to run in parallel to other system design 20× the emulated illuminance changes too rapidly for the
optimizations and guarantees accurate setpoint measurements. harvesting circuit that dynamically adapts to the environment
The exhaustive transducer characterization provided by the at a fixed interval, leading to inaccurate results. For 20× and
testbed enables informed selection of the best-performing smaller speed-ups only small deviations of less than 2.5 % are
system components. Furthermore, it facilitates system opti- observed. These high speed-up rates enable drastically reduced
mization for specific application scenarios. Examples include evaluation times, and enable a more agile design process of
scenario-specific system transducer optimization [28], iden- harvesting systems thanks to the rapid evaluation feedback.
tifying and validating transducer models and parameters, or
optimizing system parameters for calibrating maximum power C. Energy-Triggered Application Execution
point tracking (MPPT). In the following case studies the results The last case study demonstrates the use of the testbed in the
of the photovoltaic cell characterization built the foundation to end-to-end evaluation of an energy management system from
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS
Trigger Signal
Energy
Solar Panel Management Application
Unit
Environment SmartLoad
Energy Harvesting Testbed
Figure 13. The experimental setup used for the energy management
unit (EMU) evaluation. Triggered by the EMU, the SmartLoad replays
the stored current trace.
the physical signal, through internal power conditioning to the
final consumption of electrical energy. For this purpose, a so-
called energy management unit (EMU) is used to implement
the energy management principles introduced in [29]. The
EMU accumulates the energy extracted from a transducer in
a small energy buffer. This buffer, typically a small capacitor,
allows decoupling the efficient extraction of energy from sup-
plying the application, as the application is triggered no earlier
than the required energy is buffered to satisfy the requirements
of the next execution step of the application. Once enough
energy is buffered, a trigger signals the application to execute
and the buffered energy is provided as burst. Employing this
energy burst execution scheme, the EMU promises reliable and
efficient operation over a wide range of harvesting conditions
that exhibit high variability.
In this case study the EMU performance was evaluated at
different environment conditions, trigger configurations and
application profiles to understand their impact on the system
efficiency. The efficiency is defined as ratio between the energy
extracted from the transducer and the actual energy used
by the application to execute. The experiment setup used
is summarized in Figure 13: the EMU is supplied from an
AM-5412 photovoltaic cell and employs a 220 µF capacitor
as energy buffer. The integrated passive MPPT is configured
to extract energy at 75 % of the periodically sampled open
circuit voltage of the photovoltaic cell. This value represents
the optimal MPPT configuration that was extracted from the
transducer characterization in Section VII-A. The SmartLoad
replays previously stored current consumption profiles. It is
triggered by the EMU and supplied at a voltage of 2.0 V. To
understand their impact on the system efficiency the following
three parameters were adjusted for different iterations of the
experiment:
(i) the illuminance levels were varied from dim artifi-
cial light (200 lx) to bright indoor daylight conditions
(10 klx),
(ii) the energy burst size was configured to 800 µJ and
1800 µJ to understand the effect of higher or lower
trigger levels by partitioning the application into different
sized steps,
(iii) finally, the impact of consuming the energy of a burst
at different power levels was studied by consuming the
energy budget at 10 mW and 40 mW.
The observed system efficiency for different burst sizes
and power consumption levels is shown in Figure 14 as a
85
80
Efficiency [%]
75
70 800 µJ, 10 mW
65
800 µJ, 40 mW
60 1800 µJ, 10 mW
55 1800 µJ, 40 mW
50
102 103 104
Illuminance level [lx]
Figure 14. System efficiency analysis of the energy management
unit (EMU) for two different burst sizes (800 µJ and 1800 µJ) and load
power configurations (10 µW and 40 µW) at a wide range of ambient
conditions. In lower illuminance conditions, a significant advantage in
system efficiency is observed for smaller burst sizes. There is also a
marginally lower efficiency as the energy burst is consumed with higher
power.
function of the illuminance. The values represent the average
efficiency of 42 application executions at the lowest, and
up to 500 executions at the highest illuminance level. The
evaluation confirms the results presented in [29] and shows
a decreasing efficiency at low illuminance levels. With its
rapid and automated evaluation of multiple system parameters,
the testbed enables additional critical insights. The smaller
energy burst size of 800 µJ not only significantly increases
the system efficiency, but allows system to operate at even
lower illuminance levels below 200 lx (compared to 300 lx at
1800 µJ). This is due to the lower buffer voltage level of 3.5 V
compared to 4.8 V at which the application is triggered and
consequently a more efficient input voltage upconversion of
the EMU. Contrasting the efficiency for different application
power levels, a decrease is observed for the higher energy
consumption rate. However, this decrease that is attributed to
the output voltage regulation stage of the EMU is small in
comparison to the other parameters examined.
The insights gained in this case study highlight that thor-
ough energy harvesting system evaluation is only possible by
performing end-to-end system analysis. The testbed presented
enables such evaluations by jointly emulating the harvesting
environment and electrical energy consumption. It is only
through controlled and repeated experimentation of the energy
input and output that hidden internal dependencies can be
found in complex energy and power management techniques.
VIII. C ONCLUSION
An integrated testbed for emulating the physical environ-
ment conditions and dynamic energy requirements of generic
electrical loads was introduced. The emulation of the envi-
ronment’s thermal and visible light properties is combined
with the capability of sinking current traces in a time- and
event-triggered manner. The consistent reproduction of the
high dynamic range at the source and sink of energy enables
in-depth evaluation of energy harvesting systems with unprece-
dented coverage. It is complemented with a tight integration
of precision measurements and feedback control to tackle the
challenges of rapid evaluation without compromising accuracy.
IEEE TRANSACTIONS ON INDUSTRIAL ELECTRONICS

Three case studies demonstrate how the testbed excels in [17] K. Geissdoerfer, M. Chwalisz, and M. Zimmerling, “Shepherd: A
rapid and exhaustive transducer characterization, in accelerated Portable Testbed for the Batteryless IoT,” in Proceedings of the 17th
ACM Conference on Embedded Network Sensor Systems (SenSys), DOI
evaluation using real environment traces, and in analyzing 10.1145/3356250.3360042, pp. 83–95. ACM, 2019.
the impacts of diverse system parameters on the system [18] Keithley Instruments, “2450 SourceMeter® SMU Instrument -
performance in end-to-end system evaluation. Datasheet,” 2019, last accessed: 12.01.2020.
[19] W. Jiang et al., “Digitally Controlled Multi-Phase Electronic Cur-
rent Sink,” in Proceedings of the 10th IEEE International Con-
ference on Power Electronics and Drive Systems (PEDS), DOI
ACKNOWLEDGMENT 10.1109/PEDS.2013.6527138, pp. 862–866. IEEE, 2013.
The authors thank Stefan Lippuner for his inputs on the [20] M. Kazerani, “A High-Performance Controllable DC Load,” in Pro-
ceedings of the 2007 IEEE International Symposium on Industrial
circuit design aspects, Thomas Weiss and Stephan Bücheler Electronics, DOI 10.1109/ISIE.2007.4374737, pp. 1015–1020. IEEE,
for their insights into photovolatic material characterization, 2007.
Stefan Drašković for facilitating lockdown experiments, and [21] S. Upadhyay, S. Mishra, and A. Joshi, “A Wide Bandwidth Electronic
Load,” IEEE Transactions on Industrial Electronics, vol. 59, DOI
the anonymous reviewers for their feedback. 10.1109/TIE.2011.2148680, no. 2, pp. 733–739, 2012.
[22] F. Simjee and P. H. Chou, “Accurate battery lifetime estimation us-
ing high-frequency power profile emulation,” in International Sym-
R EFERENCES posium on Low Power Electronics and Design (ISLPED), DOI
10.1145/1077603.1077676. IEEE, 2005.
[1] F. Attivissimo, C. Guarnieri Calò Carducci, A. Lanzolla, and M. Spadav- [23] T. Paing, J. Shin, R. Zane, and Z. Popovic, “Resistor emulation approach
ecchia, “An Extensive Unified Thermo-Electric Module Characterization to low-power rf energy harvesting,” IEEE Transactions on Power Elec-
Method,” Sensors, vol. 16, DOI 10.3390/s16122114, no. 12, p. 2114, tronics, vol. 23, DOI 10.1109/TPEL.2008.921167, no. 3, 2008.
2016. [24] L. Sigrist et al., “Measurement and Validation of Energy Har-
[2] C. Izidoro, O. Ando Junior, J. Carmo, and L. Schaeffer, “Characteriza- vesting IoT Devices,” in Proceedings of the 2017 Design, Au-
tion of Thermoelectric Generator for Energy Harvesting,” Measurement, tomation Test in Europe Conference and Exhibition (DATE), DOI
vol. 106, DOI 10.1016/j.measurement.2016.01.010, pp. 283–290, 2017. 10.23919/DATE.2017.7927164, pp. 1159–1164. IEEE, 2017.
[3] X. Wang et al., “Extensible multi-input synchronous electronic charge [25] S. Silvestre and A. Chouder, “Effects of Shadowing on Photovoltaic
extraction circuit based on triple stack resonance for piezoelectric and Module Performance,” Progress in Photovoltaics: Research and Appli-
thermoelectric energy harvesting,” IEEE Transactions on Industrial cations, vol. 16, DOI 10.1002/pip.780, no. 2, pp. 141–149, 2008.
Electronics, DOI 10.1109/TIE.2020.3001814, 2020. [26] E. Skoplaki and J. Palyvos, “On the Temperature Dependence of Photo-
[4] B. H. Hamadani and B. Dougherty, “Solar Cell Characterization,” in voltaic Module Electrical Performance: A Review of Efficiency/Power
Semiconductor Materials for Solar Photovoltaic Cells, M. P. Paran- Correlations,” Solar Energy, vol. 83, DOI 10.1016/j.solener.2008.10.008,
thaman, W. Wong-Ng, and R. N. Bhattacharya, Eds., vol. 218, pp. 229– no. 5, pp. 614–624, 2009.
245. Springer International Publishing, 2016. [27] J. D. Kelly and J. D. Hedengren, “A Steady-State Detection (SSD)
[5] C. Domı́nguez, I. Antón, and G. Sala, “Solar Simulator for Con- Algorithm to Detect Non-Stationary Drifts in Processes,” Journal of
centrator Photovoltaic Systems,” Optics Express, vol. 16, DOI Process Control, vol. 23, DOI 10.1016/j.jprocont.2012.12.001, no. 3,
10.1364/OE.16.014894, no. 19, p. 14894, 2008. pp. 326–331, 2013.
[6] Sciencetech Inc., “(UHE-45) Large Area Solar Simulator,” [28] L. Sigrist et al., “Thermoelectric Energy Harvesting from Gradients in
http://www.sciencetech-inc.com/uhe-45-large-area-solar-simulator.html, the Earth Surface,” IEEE Transactions on Industrial Electronics, vol. 67,
2020. DOI 10.1109/TIE.2019.2952796, no. 11, pp. 9460–9470, 2020.
[7] A. M. Bazzi et al., “Solid-State Solar Simulator,” IEEE Transactions on [29] A. Gomez et al., “Dynamic Energy Burst Scaling for Transiently
Industry Applications, vol. 48, DOI 10.1109/TIA.2012.2199071, no. 4, Powered Systems,” in Proceedings of the 2016 Design, Automa-
pp. 1195–1202, 2012. tion and Test in Europe Conference and Exhibition (DATE), DOI
[8] X. Ma, S. Bader, and B. Oelmann, “Power estimation for indoor light 10.3850/9783981537079 0403, pp. 349–354. IEEE, 2016.
energy harvesting systems,” IEEE Transactions on Instrumentation and
Measurement, DOI 10.1109/TIM.2020.2984145, 2020.
[9] Chroma ATE Inc., “62050h-600s solar array simulation dc power supply
- Datasheet,” 2019, last accessed: 11.09.2020.
[10] M. C. Cavalcanti et al., “Hybrid maximum power point tracking
technique for pv modules based on a double-diode model,” IEEE
Transactions on Industrial Electronics, DOI 10.1109/TIE.2020.3009592,
2020.
[11] M. Masoudinejad et al., “A Measurement Platform for Photovoltaic
Performance Analysis in Environments with Ultra-Low Energy Har-
vesting Potential,” Sustainable Cities and Society, vol. 25, DOI
10.1016/j.scs.2015.09.006, pp. 74–81, 2016.
[12] L. Hanschke et al., “Light Insight - Emulation of Radiation Traces
for Analysis and Evaluation of Solar-Harvesting Algorithms,” in Pro-
ceedings of the 5th International Workshop on Energy Neutral Sensing
Systems (ENSsys), DOI 10.1145/3142992.3142994. ACM, 2017.
[13] S. Bobovych et al., “SunaPlayer: High-accuracy Emulation of
Solar Cells,” in Proceedings of the 14th International Confer-
ence on Information Processing in Sensor Networks (IPSN), DOI
10.1145/2737095.2737110, pp. 59–70. ACM, 2015.
[14] J. Hester, L. Sitanayah, T. Scott, and J. Sorber, “Realistic and Repeatable
Emulation of Energy Harvesting Environments,” ACM Transactions on
Sensor Networks, vol. 13, DOI 10.1145/3064839, no. 2, 2017.
[15] E. Lattanzi et al., “A Fast and Accurate Energy Source Emulator for
Wireless Sensor Networks,” EURASIP Journal on Embedded Systems,
vol. 2016, DOI 10.1186/s13639-016-0055-5, no. 1, p. 18, 2017.
[16] R. P. D. A. Filho, O. A. de Lima Junior, and C. G. F. Junior, “An
FPGA-Based Evaluation Platform for Energy Harvesting Embedded
Systems,” in Proceedings of the 32nd Symposium on Integrated Circuits
and Systems Design (SBCCI), DOI 10.1145/3338852.3339863, pp. 1–6.
ACM, 2019.