Energy Harvesting With Supercapacitors: Sizing, Cold Start and the Worst Week

On the bench, under a desk lamp, the sensor is perfect. Its small solar cell charges the supercapacitor in minutes, the radio sends a reading every few minutes, and nobody ever needs to think about a battery. Then a batch ships to a warehouse, and the installers mount them in an aisle lit by ceiling lights twelve metres up. Some nodes never send a single packet. Others send one, go quiet, send another an hour later and go quiet again. The ones near the loading-bay door work fine.
Nothing is broken. Each node is doing exactly what its energy budget allows, and nobody wrote that budget down for the aisle. Energy harvesting with a supercapacitor removes the battery, but it replaces one simple question (how many milliamp-hours?) with three harder ones. How much energy arrives on the worst day? How long must the store carry the node through the dark? And can the node start at all when the store is empty?
This post sets out a framework for answering those three questions, in that order, and then applies it to three cases: an outdoor LoRa soil sensor, an indoor warehouse node, and a node that ships empty and is installed in the dark. The numbers come from a sizing script and a one-week simulation whose inputs are stated so you can replace them with measurements from your own design.
A harvesting node is sized by its worst week, not its average day. The bench lamp is the best day it will ever see.
The Rule: Three Budgets, in Order
The framework is three budgets, checked in sequence. Each one can fail on its own, and fixing a later one never repairs an earlier one.
The energy budget comes first. Over the worst realistic period (a week of monsoon cloud, a weekend with the office lights off), the energy harvested, after conversion losses, must exceed everything the node consumes: its reports, its sleep current, the harvester chip's own quiescent draw and the store's leakage. If this budget does not close, no amount of storage helps; a larger store only delays the moment the node runs out.
The storage budget comes second. The store must carry the node through the longest gap with no harvest at all, from the moment the light goes off to the moment it returns, with margin. For a solar node outdoors that is a night plus a dark morning; indoors it may be a whole weekend.
The cold-start budget comes last. Starting from a completely empty store, the node must reach a working voltage in an acceptable time, and it must get there without falling into a loop of booting, transmitting, collapsing and booting again.
These three pull against each other. A larger store improves the second budget, costs a little on the first through extra leakage, and makes the third much slower. The design is the point where all three close together.
Budget One: Energy per Day, Worst Day First
Start with the load. For a LoRaWAN node that wakes, reads a sensor and sends a 12-byte reading every 15 minutes, the radio dominates. With 13 bytes of LoRaWAN overhead the physical payload is 25 bytes, and the airtime at 125 kHz bandwidth depends heavily on spreading factor. Our script uses the standard LoRa time-on-air formula, an assumed 40 mA transmit current at 14 dBm on a 3.3 V rail, two receive windows, 20 ms of processing and a 50 ms sensor read.
Spreading factor | Airtime per report | Energy per report | Daily load at 96 reports, including sleep |
SF7 | 61.7 ms | 11.3 mJ | 0.50 mWh |
SF10 | 411.6 ms | 59.7 mJ | 1.79 mWh |
SF12 | 1,482.8 ms | 201.1 mJ | 5.56 mWh |
Two things stand out. The spread is large: a node at SF12 needs about eleven times the daily energy of one at SF7. A harvesting design that works close to the gateway can fail at the edge of coverage, where the network server pushes it to a slower data rate. And the daily figures are small in absolute terms: half a milliwatt-hour to a few milliwatt-hours. That is why harvesting is plausible at all.
Then set the harvest against it, using the worst period, not the average. For an outdoor node that means the darkest monsoon week, measured or estimated as equivalent peak-sun hours. For an indoor node it means the dimmest location the product will be sold into, with the lights' real schedule, weekends included.
Finally, apply conversion efficiency honestly. Harvester chips quote peak efficiencies; at microwatt inputs and with a store that is low, real efficiency is lower. The script assumes 80 percent, which is optimistic at the bottom of the range.
The Standby Floor
The load table hides a term that decides most indoor designs: the power the node draws when it is doing nothing.
Three things make up the floor. The system's sleep current, including the microcontroller, the radio in its sleep state, sensor quiescent currents and any always-on regulator. The harvester chip's own quiescent consumption, which for modern parts is around a microwatt or two. And the supercapacitor's leakage, which is not zero and, as we will see, is often the largest of the three.
In the first indoor design the script models, the floor is 2 µA of sleep current at 3.3 V, 1.5 µW for the harvester chip and 2 µA of leakage at 2.7 V. That is about 13.5 µW, all day and all night. Over a week it is 2.27 mWh.
The same node indoors, with a 10 cm² cell, harvests about 2.0 mWh a week under the lighting assumed below. The floor alone exceeds the harvest. The node cannot send even one report a week on average, and no change to the reporting interval or spreading factor can fix it, because the problem is not the reports.
This is the most common reason indoor harvesting products fail. The design was checked against the energy per report and the reporting interval, and the floor was treated as negligible. At outdoor harvest levels it is. At a few microwatts of indoor light it is the whole budget.
Below a few hundred microwatts of harvest, the sleep current and the store's leakage decide the design. Measure them on the real board before choosing anything else.
What Light Is Worth
Light levels vary by more than three orders of magnitude between a warehouse aisle and a field at noon, and that range sets which designs are possible.
Outdoors, full sun delivers about 1,000 W/m², or 100 mW per square centimetre. A small crystalline cell of 12 cm² at around 12 to 15 percent efficiency gives on the order of 150 mW in full sun and a few milliwatts under heavy overcast. Even a dark monsoon day, at the equivalent of one peak-sun hour, harvests well over 100 mWh into such a cell before losses. Against a daily load of 2 mWh at SF10, the energy budget closes by a factor of 50 or more.
Indoors, the numbers collapse. Office lighting is typically a few hundred lux at desk height, corridors and warehouses often less, and light from fluorescent or LED fixtures has a spectrum quite different from sunlight. Cells optimised for indoor light, such as amorphous silicon and some newer organic and dye-sensitised types, deliver on the order of a few microwatts per square centimetre at 200 lux. The script assumes 5 µW/cm². A 10 cm² cell then gives about 50 µW while the lights are on.
The schedule matters as much as the level. Ten hours of light on weekdays and none at weekends is 50 hours a week, and the node must survive the other 118 on stored energy.
Measure rather than assume. A lux meter at the actual mounting position, at the actual angle, through a working week, is the single most useful input for an indoor design. Mounting a node on a shelf upright beneath the light instead of flat on top of it can halve the light it receives.
Budget Two: Storage Across the Longest Dark Gap
A supercapacitor stores energy as one half of capacitance times voltage squared, and only part of that is usable. The node's electronics stop working below some minimum voltage, set by the regulator and the radio. If the store works between 2.7 V and 1.8 V, the usable energy is one half of C times the difference of the squares: 2.7² minus 1.8², or about 4.05 V², times C over two.
Supercapacitor | Usable energy, 2.7 V to 1.8 V | Assumed leakage at 2.7 V | Leakage cost per day | Dark autonomy at SF10, 15-minute reports |
1 F | 2.0 J (0.56 mWh) | 2 µA | 0.13 mWh | about 7 h |
4 F | 8.1 J (2.25 mWh) | 4 µA | 0.26 mWh | about 26 h |
10 F | 20.3 J (5.63 mWh) | 10 µA | 0.65 mWh | about 55 h |
25 F | 50.6 J (14.06 mWh) | 20 µA | 1.30 mWh | about 110 h |
The leakage figures are illustrative; they vary widely by product and conditions, as the next section explains. But the pattern is real. Storage grows with capacitance, and so does leakage, roughly. A 25 F store bridges about four and a half days of darkness at SF10, but its leakage alone costs more than a 1 F store holds.
For an outdoor node the longest gap is usually a night plus a heavily overcast morning: 16 to 20 hours. A 4 F store covers that with a modest margin at SF10. For an indoor node with dark weekends, the gap is around 62 hours, and the reporting rate during that gap must fall, which is where energy-aware scheduling comes in.
Supercapacitor Behaviour That Datasheets Bury
Supercapacitors are often treated as ideal energy buckets. Four behaviours make them less simple.
Leakage is not a single number. Datasheets usually quote leakage current after the part has been held at rated voltage for a long period, often 72 hours. In the first hours after charging, the apparent leakage is much higher, because charge is still redistributing into the electrode structure. Leakage also rises with voltage and with temperature. A design that holds the store at its maximum voltage on a warm afternoon pays the highest leakage at exactly that moment. Some designs deliberately cap the charge voltage below the maximum for this reason.
ESR sets the transmit dip. When the radio draws 40 mA, the store's voltage drops instantly by the current times the equivalent series resistance. Large cylindrical supercapacitors have ESR well under an ohm. Small coin-type parts can have tens of ohms; at 30 Ω, a 40 mA burst drops the voltage by 1.2 V, enough to brown out the node even when the store holds plenty of energy. A low-ESR ceramic or polymer capacitor in parallel, sized to supply the transmit burst, solves this.
Voltage per cell is limited. A single electric double-layer cell is typically rated around 2.7 V. Two in series reach 5.4 V but need balancing, because mismatched leakage lets one cell drift over its rating. Lithium-ion capacitors and other hybrid parts offer higher voltage and lower leakage, with their own charge and temperature rules.
Life depends on temperature and voltage. Like electrolytic capacitors, supercapacitors age faster when hot and when held at full voltage. An outdoor enclosure in Indian summer sun can run well above the air temperature, which belongs in the life estimate.
Read the leakage figure's test condition before you use it. A leakage current measured after 72 hours at 25 °C says little about the first night after a sunny afternoon at 50 °C.
Budget Three: Cold Start From Empty
A node with an empty store faces a bootstrap problem. Its main boost converter needs power to run, and the power has to come from the harvester. Harvester chips solve this with a separate cold-start circuit, usually a low-efficiency charge pump, that runs directly from the source until the store reaches a voltage where the main converter can take over.
Cold-start specifications are worth reading closely. One widely used harvester chip quotes a cold-start input of 330 mV typical but 600 mV guaranteed. Another quotes cold start from 380 mV and 3 µW of input power. A small solar cell under dim light may sit right at those thresholds, so a node can harvest normally once running but fail to cold-start in the same light.
Even after cold start, filling the store takes time. The script estimates the time to bring a store from empty to 1.8 V at 80 percent efficiency, ignoring leakage:
Source | 1 F store | 10 F store |
Indoor cell, 10 cm² at 200 lux (about 50 µW) | 11.3 h | 112.5 h |
Indoor cell, 10 cm² at 1,000 lux (about 300 µW) | 1.9 h | 18.8 h |
Outdoor cell, 12 cm², overcast (about 5 mW) | 7 min | 68 min |
Outdoor cell, 12 cm², sun (about 150 mW) | under 1 min | about 2 min |
Outdoors, cold start is a non-issue. Indoors, a 10 F store at 200 lux takes more than four days of continuous light to reach operating voltage, and with lights off at night and leakage running, it may never get there. A node installed on a Friday afternoon may not send its first packet until the middle of the following week, by which time the installer has filed it as faulty.
The trade-off is direct: the store that gives the best dark autonomy gives the slowest cold start.
The Boot Loop and How to Avoid It
The second cold-start failure is worse than slowness. When the store reaches the voltage where the microcontroller can run, the firmware boots, initialises the radio and transmits. The transmit burst pulls the store below the brown-out threshold, the microcontroller resets, the store recovers slightly, and the cycle repeats. Each cycle spends energy on a boot and a partial transmission, so the store never accumulates enough to escape. From outside, the node appears dead or sends a corrupted packet every so often, which is exactly the warehouse behaviour from the opening.
Two mechanisms prevent it, and good designs use both.
The hardware gate. Most harvester chips provide a "storage OK" output with programmable thresholds and hysteresis. Use it to hold the load switch or the regulator enable off until the store is well above the level a full transmit cycle needs, and keep it on until the store falls to a lower threshold. The hysteresis gap is the energy reserve for at least one complete wake, transmit and receive cycle.
The firmware gate. On every boot, before touching the radio, read the store voltage and compare it with the energy a transmission needs. If it is too low, go straight back to the deepest sleep available for a fixed time, without initialising anything expensive.
#define V_TX_MIN_MV 2350 /* enough for one full uplink + RX windows, with ESR dip */
#define V_RECHECK_S 600
void on_boot(void) {
uint16_t vcap = adc_read_vcap_mv(); /* divider switched in only for the read */
if (vcap < V_TX_MIN_MV) {
deep_sleep_seconds(V_RECHECK_S); /* radio never initialised */
return;
}
radio_init();
if (!lorawan_joined()) lorawan_join(); /* joins are expensive: retry slowly */
schedule_next_report(vcap);
}
The threshold is not a guess: it is the minimum voltage, after subtracting the ESR dip at transmit current, from which one complete uplink cycle at the slowest expected spreading factor still ends above the brown-out level.
Energy-Aware Scheduling
The firmware gate prevents collapse. Energy-aware scheduling keeps the node useful across changing light. The idea is simple: the reporting interval follows the stored energy, so the node reports often when the store is full and slows down as it drains.
uint32_t next_interval_s(uint16_t vcap_mv) {
if (vcap_mv >= 2550) return 300; /* store nearly full: every 5 minutes */
if (vcap_mv >= 2350) return 900; /* normal: every 15 minutes */
if (vcap_mv >= 2150) return 3600; /* conserving: hourly */
return 0; /* survival: no uplinks, recheck later */
}
To test the idea, we simulated one week, second by second, for an indoor redesign described in the next sections: a 30 cm² cell at 5 µW/cm² for ten hours on weekdays and dark weekends, a 5.5 µW floor, SF7 reports of 11.3 mJ, and a 1.8 V cut-off. The adaptive schedule above was compared with a fixed report every nine minutes, which the weekly energy balance would appear to allow.
Store | Adaptive schedule | Fixed 9-minute schedule |
1.0 F | 875 reports, no failed attempts, lowest voltage 1.85 V | 901 attempts succeeded, 219 failed, store ran below the 1.8 V cut-off |
1.5 F | 1,017 reports, no failures, lowest 2.04 V | 972 succeeded, 148 failed, ran below cut-off |
3.0 F | 1,242 reports, no failures, lowest 2.24 V | 1,120 succeeded, no failures, lowest 1.96 V |
The fixed schedule spends the weekend's energy on Saturday and is dead by Sunday; the adaptive one slows down and is still reporting hourly on Monday morning. With a 3 F store, the adaptive schedule delivers more reports than the fixed one and never approaches the cut-off.
Case One: An Outdoor LoRa Soil Sensor
Apply the framework to a soil moisture node in an open field, reporting every 15 minutes, with a 12 cm² crystalline cell.
Energy budget: at SF10 the node needs about 1.8 mWh a day plus leakage. A dark monsoon day harvests well over 100 mWh before losses. The budget closes by a wide margin even at SF12. Outdoors, energy is rarely the constraint.
Storage budget: the longest regular gap is a night plus a dark morning, around 18 hours. A 4 F store gives about 26 hours of autonomy at SF10, which covers it. A run of very dark days is the edge case; the energy-aware schedule handles it by slowing to hourly reports when the store drops, which a farmer will rarely notice.
Cold-start budget: minutes in any daylight. Not a concern.
The design risks outdoors are elsewhere. Heat shortens supercapacitor life, so the enclosure needs shading and the store should not sit at full voltage on hot afternoons. Dust on the cell reduces harvest over a dry season; a cell mounted at a steep angle sheds dust and rain better. And the node must survive the occasional total failure gracefully: after a week of darkness it should cold-start and rejoin the network on its own, which the boot gate makes possible.
For this case, a battery-free design is attractive: no replacement visits, a long service life and no lithium cells to transport. Our earlier post on solar and battery sizing for pond and shed nodes covers the battery alternative for higher-power nodes.
Case Two: An Indoor Warehouse Node
The warehouse node from the opening is the harder case, and the framework shows exactly why the first design failed.
Energy budget, first design: 10 cm² of indoor cell at about 200 lux, 50 µW for ten hours on weekdays, gives about 2.0 mWh a week after conversion losses. The floor (2 µA sleep, 1.5 µW harvester chip, 2 µA leakage) is 2.27 mWh a week. The budget fails before a single report is counted. That is the "never sends" group from the opening.
Redesign: three changes close the budget. The cell grows to 30 cm², tripling the harvest to about 6.0 mWh a week. Sleep current falls to 0.8 µA through a lower-power sensor and regulator, and the store changes to a smaller, lower-leakage part at about 0.5 µA. The floor falls to about 0.92 mWh a week. After a 30 percent margin, the remaining energy supports a little over 1,100 SF7 reports a week, about one every nine minutes on average.
Storage budget: the weekend is 62 hours without light. The floor alone costs about 0.34 mWh over that gap, and hourly reports at SF7 add about 0.2 mWh. A 1 F store holds 0.56 mWh usable, which leaves almost no margin. The simulation confirms that 1.5 to 3 F with energy-aware scheduling carries the node through the weekend with voltage to spare.
Spreading factor: indoors the gateway must be close enough for SF7. A warehouse node pushed to SF10 needs about five times the energy per report, and the budget would close only at a much lower reporting rate.
Case Three: Shipped Empty, Installed in the Dark
The third case is about logistics rather than physics. Supercapacitors are often shipped discharged, and even a charged one leaks away over weeks in a warehouse. So every node starts life with an empty store, and its first hours happen wherever the installer puts it.
With the redesigned 30 cm² cell at 200 lux, a 1.5 F store needs about ten hours of light to climb from empty to the 2.35 V transmit threshold: a full working day of office lighting. With nights, leakage and a dim corner, the first uplink slips into the following week. An installer who mounts twenty nodes on Friday and checks the dashboard on Monday sees most of them missing.
Three remedies work together. A commissioning step: before mounting, each node is placed under strong light (a work lamp or daylight at a window) until its status LED confirms it has joined the network, which takes minutes. A small primary cell as a starter: some designs add a tiny lithium primary cell used only to cold-start the node and to carry it through exceptional darkness, with the harvester doing the daily work. And a smaller store, accepting shorter dark autonomy in exchange for faster cold start, where the lighting schedule allows.
A new harvesting node has no energy and no history. Give it light and a network join before it leaves the installer's hands.
Whichever is chosen, write the installation procedure around it. The best harvesting design still fails in the field if the installer does not know that a new node needs light before it can speak.
When a Battery Is the Better Answer
Harvesting with a supercapacitor is not always the right choice. The framework itself shows where it loses.
High reporting rates or heavy payloads at long range, such as frequent SF12 uplinks or GNSS fixes, can exceed what small cells harvest in poor light, and a battery's energy density makes the product smaller and cheaper. Very dark sites, such as plant rooms, basements and enclosed cabinets, may harvest too little to close even the floor. Products that must work immediately out of the box, with no commissioning step, are easier with a primary cell. And very cold sites reduce both harvest and store performance.
Hybrid designs often win. A rechargeable lithium cell charged by the harvester gives far more storage than a supercapacitor of the same size, at the cost of charge-temperature limits and a finite cycle life. A primary cell with a harvester extending its life, rather than replacing it, removes the cold-start problem entirely and still stretches service intervals.
The honest comparison is total cost over the product's life: battery replacement visits, their labour and travel, against the extra cell area, the lower-leakage store and the commissioning step that harvesting needs. For nodes in fields, on roofs and on structures where every visit is expensive, harvesting often wins. For nodes on a shelf that a technician passes every week, a ten-year primary cell may be simpler.
Testing a Harvesting Node
Harvesting nodes need tests that ordinary battery products do not.
Light at known levels. A light box or a dimmable fixture with a lux meter at the cell's position lets you measure harvest at 100, 200, 500 and 1,000 lux on the production cell and harvester chip, rather than relying on the cell's datasheet.
The floor, measured. A source-measure unit or a low-current meter on the whole node in sleep, with the store at its normal operating voltage, gives the number that decides indoor designs. Measure store leakage separately, at several voltages and temperatures, starting soon after charging and again after several days.
Cold start from 0 V. Short the store, remove it from light for a day, then expose it to the dimmest specified light and time the first successful uplink. Repeat at the lowest operating temperature.
Long dark. Charge to full, switch off the light for the longest specified gap, and confirm the node is still reporting at the end, at the reduced rate the schedule allows.
Boot-loop hunting. Ramp the light slowly up from darkness and watch for repeated resets near the threshold. If the reset counter climbs, the hardware or firmware gate needs more hysteresis.
A Monday-Morning Sequence
For a team starting a harvesting node, this order avoids most of the failures above.
Define the worst period. Measure lux at real mounting positions through a working week, or obtain monsoon-season irradiance for outdoor sites.
Measure the floor first. Sleep current of the full board, harvester chip quiescent draw, and store leakage at operating voltage and temperature.
Close the energy budget. Worst-period harvest after realistic conversion losses must exceed floor plus reports, with margin, at the slowest spreading factor the node will see.
Size the store for the longest dark gap. Include leakage at the operating voltage, and plan reduced reporting during the gap.
Check cold start at the dimmest light. Confirm the harvester chip's guaranteed threshold, not its typical one, and time the first uplink from 0 V.
Add both gates. Hardware hysteresis on the storage-OK output, and a firmware voltage check before any radio activity.
Add energy-aware scheduling. Map store voltage to reporting interval, and test it across a simulated week.
Handle the transmit dip. Measure the store's ESR and add a low-ESR buffer if the dip approaches the brown-out level.
Write commissioning into the installation procedure. Every new node gets light until it has joined.
Back to the Warehouse Aisle
The nodes that never sent a packet were the ones whose floor exceeded their harvest. The ones that sent one packet and went quiet were caught in a boot loop or ran dry on a fixed schedule. The ones by the loading-bay door had daylight, and daylight hid every mistake.
None of these failures needed new technology to fix: a larger cell, a lower floor, a store sized for the weekend rather than the night, a voltage gate in firmware and a commissioning step for the installers. The difference between a harvesting product that works and one that does not is whether someone wrote down all three budgets before the batch shipped.
Related service: End-to-End IoT Product Development
Related reading: Sizing Solar and Battery for Pond and Shed IoT Nodes in India, Battery Management for IoT Devices: Protection, Charging and Fuel Gauging Without a Full BMS and LoRaWAN FUOTA: Airtime, Fragments and What a Firmware Update Really Costs a Fleet
EurthTech is an engineering-first product design and embedded systems company with a development center in Hyderabad. To discuss a project, write to connect@eurthtech.com or see https://www.eurthtech.com/case-studies.




Comments