Energy Harvesting Electronics: Cold Start, Storage Leakage, and What the Power Budget Actually Funds
Quick Overview
Problem: A harvester that meets its specification on the bench powers an installed fleet that does not hold its reporting rate.
Common failure points: Budgets built from average harvested power, storage leakage omitted from the power tree, cold-start thresholds validated from a charged capacitor, and transmission energy underestimated at architecture time.
Where it appears: Indoor building-automation sensors, industrial condition-monitoring nodes, battery-free switches and asset tags, agricultural and environmental LPWAN deployments.
Engineering focus: Worst-interval source characterization, PMIC cold-start and hysteresis design, leakage-inclusive energy budgeting, energy-aware firmware scheduling.
Energy harvesting programs tend to run into trouble in a specific place. The transducer meets its specification, the power management IC operates, the firmware runs — and the installed fleet still falls short of the reporting rate the product was sold on.
The usual cause is that the energy budget was built by comparing two averages: average harvested power against average load power. That comparison is adequate for a mains-fed or battery-fed device. For a harvested one it omits both of the quantities that govern the design — the worst sustained interval the source will produce, and the peak current the load demands for a few milliseconds at a time. Between them sit the storage element and its leakage, the converter’s cold-start threshold, and the hysteresis window that decides when the load is permitted to run.
What follows works that budget through with the assumptions written out, then lists what to measure on your own design. The numeric examples in section 8 are calculations from published component and material data. They are illustrative, and they are not measurements from a particular client program; where Promwad project experience is referenced, it is referenced at the scope already published.
Wrong Assumption
The assumption behind most early budgets is that average harvested power exceeding average load power implies indefinite operation. On a harvested node, neither term in that comparison is stable. Harvested power varies by an order of magnitude across a day and again across a season. Load power is dominated by short radio bursts that the source cannot supply directly and that the storage element has to cover. A budget written as two averages has no way to express either variation, which is why it can pass review and still be wrong by a wide margin in service.
Why It Fails
The worst interval sets the storage requirement. Published characterization of commercial indoor-optimized amorphous-silicon modules puts output at roughly 5–9 µW/cm² at 200 lux and around 46 µW/cm² at 1000 lux, with incident optical power in the range of 55–300 µW/cm² across that span. A bench lit to 1000 lux therefore overstates a corridor at 150 lux by close to an order of magnitude, and neither figure describes a holiday weekend with the lighting switched off. The storage element has to carry the longest dark period the installation will see, and that period is a site measurement.
Storage leakage is a continuous load. Reported self-discharge for conventional supercapacitors runs roughly 10 to 50 times higher than for a rechargeable lithium microbattery. In one published measurement on an EnOcean-based sensor module, supercapacitor leakage averaged about 5 µA and accounted for approximately 40% of the module’s total deep-sleep current. On a microwatt harvest that is a significant share of the daily yield, and it appears in no side-by-side capacity comparison. Lithium-ion capacitors sit between the two technologies with markedly better self-discharge than symmetric EDLCs.
Cold start and run-on are separate thresholds. Texas Instruments specifies the bq25570 cold-start input at VIN ≥ 330 mV typical from a depleted storage element, with continued harvesting down to VIN = 100 mV once the boost converter is running, at a full operating quiescent current of 488 nA. The e-peas AEM10941 is specified to cold-start from 380 mV and about 3 µW, then harvest from 50 mV. A dim mounting position can sit above the run-on level and below the cold-start level for a long time. Validating from a charged capacitor will never show this, which makes cold-start-from-empty a distinct acceptance case for MCU firmware engineering on this class of device.
Transmission energy dominates everything else. Firmware effort concentrates on sleep current because that is what a profiler shows over a long capture. On a harvested node the radio bursts carry the budget. A BLE advertising event costs on the order of tens of microjoules; a LoRaWAN uplink at SF12 costs on the order of hundreds of millijoules — a spread of roughly four orders of magnitude. Wireless technology selection is therefore an energy decision taken at architecture time, which is the frame the trade-offs in low-power wireless technology selection are best read in.
Transducer coupling is part of the specification. A thermoelectric module develops voltage across the temperature difference that falls on it, which is a fraction of the source-to-ambient difference once the interface material and heat sink have taken their share. Because output power scales with the square of the module ΔT, a modest error in the thermal path produces a large error in power. Resonant vibration harvesters have the mechanical analogue: output is concentrated in a narrow band around the resonant frequency, industrial vibration commonly sits in the 50–100 Hz range, and most MEMS harvesters resonate well above it.
These interact. A node that cold-starts late in the morning has already lost the brightest hours, which shortens the window in which it can both run and recharge. A radio schedule sized against summer illuminance will be short of energy in November. A storage element chosen on cycle life can leak away the margin that would have covered either.
Hidden System Complexity
ambient source → transducer → impedance match / MPPT → rectifier and boost converter → storage element → leakage → regulation → MCU sleep current → radio burst → network schedule → server expectation
A gap in a server-side time series commonly originates several stages upstream. A shadowed or soiled cell shifts the operating point, the converter runs at lower efficiency, the storage element never reaches the PMIC’s VBAT_OK threshold, and the load is never enabled — with no error raised anywhere in the chain.
Maximum power point tracking carries an assumption worth checking. Both reference PMICs above track by fractional open-circuit voltage: the converter periodically disconnects the source, samples VOC, and regulates the input to a configured fraction of it. That fraction is set in hardware and is typically around 70–80% for photovoltaic sources and near 50% for thermoelectric ones. A late change of cell technology without re-deriving the ratio leaves the harvester operating off its maximum power point while every component in the BOM still matches its datasheet.
The compute class underneath fixes both the sleep current and the cost of each wake, which places this decision alongside the ones discussed in choosing an edge device architecture for constrained environments.
Failure Patterns
Three patterns recur across this class of deployment. They are described here by mechanism, since the observable symptom is usually the same in each: a node that stops reporting without raising a fault.
The commissioning loop. A battery-free node installed in a dim space charges to the point where its load is enabled, boots, immediately attempts a network join, collapses the rail with the radio current, browns out, and repeats. The device is cycling rather than dead. Each cycle consumes charge that would otherwise accumulate, so the loop can persist for as long as the source holds. The usual root cause is a hysteresis window narrower than the energy required to complete one transmission.
Seasonal dropout. An indoor photovoltaic deployment commissioned and validated in summer loses reporting rate through autumn, concentrated in north-facing rooms and in corridors with occupancy-controlled lighting. No firmware changed. The illuminance floor at those positions was never characterized, and the storage element was sized for a night instead of a shutdown period.
Output loss after a process or maintenance change. A resonant vibration harvester delivers specified power for an extended period, then drops sharply after a bearing replacement, a speed change, or a new duty profile. The harvester is intact and correctly mounted; the dominant excitation frequency has moved off its resonance band. The same class of problem appears thermally when a process setpoint or insulation change alters the available ΔT.
Low-Power and Energy Harvesting Engineering
Harvested-node problems — commissioning loops, seasonal dropout, leakage-dominated budgets, reporting rates the source cannot fund — are addressed at the level of source characterization, power architecture, and firmware scheduling. Promwad designs low-power hardware and firmware for battery-powered and self-powered embedded products, including ultra-low-power MCU and regulator selection, power-tree and wake/sleep architecture, system-level power profiling against real usage scenarios, and energy harvesting integration planning for solar, kinetic, and RF sources.
Engineering Experience Across Low-Power Silicon and Power Management Platforms
Worked Example — An Indoor Photovoltaic Node Budget
This section is a calculation, with the assumptions listed. Inputs are taken from published indoor-PV characterization, published PMIC datasheets, and published supercapacitor leakage measurements. Substitute your own measured values before using any of it as a design input.
Assumptions: a 10 cm² indoor-optimized amorphous-silicon module at approximately 200 lux for ten hours per day, taken at the low end of published output (6 µW/cm²); 80% PMIC conversion efficiency; 2 µA storage leakage and 1.5 µA MCU deep-sleep current, both at 3.3 V and both continuous across 24 hours.
Budget line Daily energy Running total Raw harvest (60 µW × 10 h) 2.16 J 2.16 J PMIC conversion loss at 80% −0.43 J 1.73 J Storage leakage, 2 µA at 3.3 V, 24 h −0.57 J 1.16 J MCU deep sleep, 1.5 µA at 3.3 V, 24 h −0.43 J 0.73 J Available for sensing and transmission 0.73 J 34% of raw harvest Two overheads that appear in no capacity comparison — leakage and sleep current — take about half the harvest before any work is done. What the remainder buys, at representative transmission costs:
Transmission Approx. energy per event Events per day on 0.73 J BLE advertising event ~25 µJ ~29,000 BLE connection event ~150 µJ ~4,900 LoRaWAN uplink, SF7 ~25 mJ ~29 LoRaWAN uplink, SF10 ~150 mJ ~5 LoRaWAN uplink, SF12 ~600 mJ ~1 A requirement written as “a report every fifteen minutes over LoRaWAN” asks for 96 uplinks a day. Against this budget at SF10 that is short by roughly a factor of twenty. The figure moves with cell area, illuminance, leakage, and spreading factor, so it is worth recomputing with site values — but the order of magnitude is what makes it an architecture-phase question.
A second calculation, on the thermal side. Take a surface at 60 °C and ambient at 25 °C, so 35 K is available. Assume a module thermal resistance of 5 K/W, an interface of 2 K/W, and a compact heat sink at 20 K/W. The thermal divider puts about 6.5 K across the module. Since output scales with ΔT², the power is roughly 3% of what a 35 K assumption would predict. Taking the external resistance down toward the module’s own — in this example a sink at 5 K/W rather than 20 — raises the module ΔT to about 14.6 K and the power to roughly 17% of the ideal figure. The sink geometry, in other words, is a power specification, and it is decided by the enclosure.
Solution Approach
Step 1: Characterize the source where the device will be mounted. Log illuminance, surface and ambient temperature, or vibration spectrum at the actual mounting positions across a full duty cycle, including nights, weekends, and the seasonal minimum where it applies. The output is a distribution with a floor: the worst sustained level and how long it lasts. This is the one input that cannot be taken from a datasheet, and everything downstream depends on it.
Step 2: Build the budget in order, leakage first. Raw harvest, conversion loss, storage leakage, regulator and MCU quiescent current, and only then the energy left for sensing and transmission — the sequence used in section 8. Measure the leakage of the specific storage part at the operating voltage and temperature; published figures vary widely between technologies and between samples.
Step 3: Treat cold start as its own acceptance case. From a fully depleted storage element, at the worst-case source level, measure the time to first successful transmission and record it as a test result. Size the hysteresis window so the system accumulates enough charge to complete a full transmission before the load is enabled. The commissioning loop described above is reproducible on the bench as soon as the test starts from an empty capacitor.
Step 4: Schedule against stored energy. A fixed timer cannot represent a variable source. Read the storage voltage, convert it to available energy, and decide from that: skip, aggregate, reduce the data rate, or transmit. Where a transaction may not complete within one energy window, the system needs checkpointing into non-volatile state so a brownout costs progress instead of restarting the operation. On an RTOS this belongs in the power-management subsystem, which is one reason Zephyr RTOS-based firmware is a reasonable default here.
Real Trade-offs
Supercapacitor against rechargeable lithium microbattery. Supercapacitors accept charge quickly, tolerate high cycle counts and cold temperatures, and need no charge-management circuitry, at the cost of leakage that is material on a microwatt budget. Lithium microbatteries and lithium-ion capacitors hold charge far better and bring chemistry limits, charge control, and finite cycle life. Hybrid storage — a capacitor for the burst and a cell for the reserve — resolves much of it, provided the interface between them does not add quiescent current that cancels the gain.
Fractional-VOC tracking against a fixed operating point. Fractional open-circuit tracking costs almost nothing in quiescent current and works well when the source matches its configured ratio. A fixed operating point is simpler and more predictable, and gives up yield when illumination or ΔT varies widely. Because the ratio is set in hardware, a late source change makes it silently wrong.
Cell area against enclosure and installation. Harvest scales linearly with area, which makes a larger cell the most dependable lever available. It also collides with industrial design, mounting position, and unit cost, and a cell that must stay clean and unshaded transfers a maintenance obligation to the customer — often the same obligation the battery-free proposition was meant to remove.
Fully battery-free against harvester-assisted. Removing the battery eliminates replacement logistics, disposal obligations, and a dominant line in the use-phase footprint, which is the accounting discussed in eco-design for embedded hardware. A primary cell backing a harvester gives deterministic behavior during dark periods and extends the replacement interval substantially, at the cost of the battery-free claim. The useful form of this decision is a service-level statement: what reporting rate is guaranteed in the worst interval, with the cell and without it.
Spreading factor against energy per report. On LPWAN, range and energy trade directly, with airtime and energy roughly doubling per step in spreading factor. Placing gateways so nodes can operate at a lower spreading factor is frequently cheaper than enlarging every harvester in a fleet, which makes it a network-planning decision as much as a device one — the scale at which LPWAN sensor network design is usually argued.
Over-the-air updates. A firmware image is the largest single energy transaction such a device will perform, and it has to survive power loss mid-transfer. Delta updates, resumable transfers, and A/B partitioning each cost flash and complexity; without them an update campaign can strand a fleet that has no technician route. The provisioning and update obligations set out in embedded lifecycle management need to be sized against the energy budget at design time.
Typical Energy Harvesting Engineering Tasks
Source Characterization and Energy Budgeting
Measuring illuminance, thermal, or vibration profiles at real mounting positions across a full duty cycle, and deriving the worst-interval budget the design has to hold.
Power Architecture and PMIC Design
Harvester-to-storage topology, cold-start and hysteresis thresholds, MPPT configuration, storage technology selection, and leakage budgeting across the power tree.
Energy-Aware Firmware Development
Wake/sleep state machines driven by stored energy, checkpointing for intermittent operation, adaptive transmission policy, and power-managed RTOS integration.
Power Profiling and Field Validation
Peak-current and duty-cycle measurement on target hardware, cold-start-from-empty acceptance testing, and long-duration logging against real usage profiles.
Qualifying Symptoms
- Devices operate on the bench and fail to join the network at their mounting position.
- Fleet reporting rate tracks the season or the building’s lighting schedule.
- Nodes cold-start each morning, indicating the storage element did not carry the dark period.
- A harvester meets its specified power in isolation and falls short once mounted in the enclosure.
- Vibration or thermal harvester output drops after maintenance or a process change, with no fault indicated.
- The energy budget states an average harvested power with no worst-case interval or dark-period reserve.
- Transmission intervals are set by a timer, with no measurement of stored energy before the radio is enabled.
- The storage element was selected on capacity and cycle life, with leakage current absent from the power budget.
Solution Context Link
At this stage the work sits in power architecture and energy budgeting: a measured source distribution with a stated floor, a leakage-inclusive daily budget, a cold-start and hysteresis design validated from an empty storage element, and firmware that reads stored energy before enabling the radio.
Where the device is an instrument as well as a radio, the analog front end and acquisition chain carry their own budget, and industrial sensor development is where sampling strategy and signal conditioning get designed against available energy. For building deployments, the commissioning and network constraints in building automation systems engineering often determine the reporting rate before the harvester does. Where the budget is being reconstructed on a product that already shipped, battery life optimization for IoT devices sets out the same method applied to a cell.
Promwad Engineering Experience
The gap this article is about — a controlled measurement that does not survive an installed condition — is one Promwad has measured and published on battery-powered devices. In published work on low-power IoT sensor and wearable electronics, field battery life landed substantially below bench measurement on identical hardware and firmware, and the article setting out why is why wearable IoT sensor electronics fail at battery life in field conditions. The same mechanism governs harvested designs, with a smaller margin to absorb it.
On the hardware side, published project scope relevant to this class of device includes firmware porting with power-consumption optimization for solar activity sensors, low-power three-phase metering on an STM32L4 platform with NB-IoT and LoRa connectivity, and ultra-low-power wearable platforms combining BLE with LPWA. These are listed with their documented scope in the case links below.
FAQ
How much power can realistically be harvested indoors?
Supercapacitor or rechargeable battery for storage?
Why does a node fail to start but run normally once started?
Can a harvested node run LoRaWAN?
Will 3GPP Ambient IoT replace purpose-built harvesting designs?
Related Engineering Cases
- Porting Solar Activity Sensor MCU Firmware to New PCB: ARM MCU firmware port with BSP development and power-consumption optimization against client requirements; LIN conformance verified in an external laboratory.
- Three-Phase Power Meter Sensor Development: Low-power metering design on STM32L4 with STPM33/STPM34 analog front end, ST7580 power line communication, and NB-IoT and LoRa connectivity.
- Firmware Development for a Connected Bicycle Computer: Low-power embedded firmware with BLE connectivity and sensor integration for wearable sports hardware.
- Health Monitoring Wearable Ecosystem on Nordic nRF5340: Wearable platform combining BLE and LPWA Cat M1, built on a low-power SoC selected for battery life alongside sensor-processing capability.
- Predictive Edge-AI Monitoring for Ventilation Systems: On-device condition monitoring that transmits compact health indicators instead of raw sensor data, which is the transmission-energy strategy a harvested node depends on.
- Electronic Unit for a Wireless Charging Hub: Wireless power receiver design to Qi power class 0 with foreign-object detection and charge control.