IoT Battery Life Calculator: Sleep, Wake, Transmit & Self-Discharge

The load model integrates five sequential states over one start-to-start reporting interval. Every state current is the total node current in the selected measurement domain. The unoccupied time is sleep.

For battery-terminal or direct/LDO currents, $g=1$. For regulated-rail currents through a switching converter, $g=V_{rail}/(V_{bat}\eta)$. The separate board and regulator currents are always battery-side additions. Power conversion losses are allocated to the state that causes them. The daily-energy chart excludes internal self-discharge.

Available charge starts at $Q_0=Q_{nom}f_Tf_{cut}f_{pulse}$. These three factors come from the user's cell data and are not universal temperature or pulse correlations.

The annual self-discharge fraction $d$ is the selected starting fraction times the multiplier. We adopt a first-order remaining-charge approximation: $k=-\ln(1-d)/8766$ per hour, using a Julian year. At a constant mean load, $dQ/dt=-I_{avg}-kQ$. Its solution is

At zero self-discharge, $Q(t)=Q_0-I_{avg}t$ and $L=Q_0/I_{avg}$. Numerically stable log1p and expm1 functions preserve this limit. The animation stops at $L$ and clamps remaining charge to zero.

Charge accounting is derived here from conservation, following the mode-weighted budgeting approach in Silicon Labs, section 3, and Callebaut et al., pp. 4–5 and 17–18. Published reference checks use JLCPCB's custom/development-board examples and Callebaut Table 7, p.30. The waveform diagram uses schematic state widths and heights; numerical durations and currents are labeled.

Repeated-cycle applicability. The lifetime is a continuous average-current approximation. The estimated cycle count is $N=L/T$, and $Q_0/q_{cycle}$ compares available charge with one cycle's load demand. Lifetime results are withheld if the battery cannot fund a full cycle or the averaged lifetime is shorter than one interval. Below ten estimated cycles a warning calls for state-by-state analysis; ten is a practical averaging threshold, not a cell qualification rule.

The interval sweep recomputes nonnegative sleep time at every point. A zero-load point has unbounded depletion time in this model and is plotted as an intentional gap in both lifetime curves. This also applies with first-order self-discharge alone, which approaches zero charge asymptotically. Every strictly positive-load sweep point uses the documented lifetime equation.

• The defaults combine ESP32-like currents with illustrative timing and low board leakage. Espressif's ESP32 Series Datasheet v5.3 identifies a 10 µA chip deep-sleep configuration. The other default active-state values and the 3 µA board / 2 µA regulator additions are design scenarios; replace them with measurements.
• Battery capacity and nominal voltage starting points come from the named manufacturer documents. CR2032: 235 mAh to 2.0 V at 21 °C and a 15 kΩ load. Saft LS 14500 September 2023: 2.6 Ah at 14 mA, 20 °C, 2.0 V cut-off. VARTA 4906 June 2018: 2970 mAh under high-resistance discharge; the two-cell series pack is 3 V, 2970 mAh. Panasonic NCR-18650BF: minimum nominal capacity 3250 mAh, 3.6 V.
• Retained-capacity defaults 0.90, 0.90 and 0.95 are independent design allowances. They are not fitted to those datasheets. If capacity already reflects the same temperature, cut-off or pulse test, set the corresponding factor to 1.
• Annual-loss starting fractions are 0.01 for primary lithium, 0.025 for alkaline and 0.12 for Li-ion. Energizer and Saft support approximately 1% or less at reference storage conditions; alkaline and Li-ion values are generic starting values from Callebaut Table 3, p.17. Verify the actual cell. With a custom battery, the selected row still supplies the annual starting fraction; adjust its multiplier.
• A single mean battery voltage and switching efficiency apply across all states. Direct/LDO current transfer assumes adequate voltage headroom throughout the usable charge window.
• Currents are constant within each state, cycles repeat continuously, and self-discharge acts on remaining available charge. Published storage losses quoted relative to nominal capacity are approximated by this first-order model. Cell aging, protection-circuit current and environmental leakage require separate allowances or board-current entries.
• The largest modeled state current is an average within that state; it can be lower than the real instantaneous peak. Coin-cell and Li-SOCl2 pulse conditions are reminders based on limited datasheet examples, not battery qualification tests.
• Averaging assumes many repeats of the same cycle. Fewer than ten predicted cycles triggers a warning; less than one supported cycle blocks lifetime results. Starting phase is unspecified, so this tool does not resolve within-cycle depletion.

Treat the output as an estimate and measure the real device at its battery terminals. Capture enough full cycles to include startup, retries and variable network attachment. Development-board current can be orders of magnitude above the chip's sleep specification; SparkFun measured approximately 2.5 mA total deep sleep on its ESP32 Thing Plus USB-C board in one configuration.

Use positive available capacity and mean current, an annual self-discharge fraction below 1, and a reporting interval at least as long as the active states. Zero-duration states are omitted from the budget. The annual fraction is used without clipping throughout 0 ≤ d < 1; invalid fractions block results. A current measured for the whole board should not be counted again as separate regulator or board overhead.

The model estimates charge depletion, without predicting terminal voltage, brownout, regulator dropout, voltage recovery, cold-start behavior, Li-SOCl2 passivation, pulse-buffer sizing, a Peukert effect, recharge cycles or harvesting. A capacity derating factor cannot make an unsupported pulse feasible. Check the exact manufacturer's curves at your cut-off voltage, temperature, load and storage history. Switching-converter efficiency at microamp sleep loads may differ substantially from its active-load efficiency.

Long projections can exceed practical calendar life even when the algebra is valid. Independently validate any service-life commitment. The interval chart is a sensitivity study, including long projections; it is not a promise of field life. The animation is an averaged charge inventory rather than a voltage discharge curve.

Source limitations: Adafruit's profiling guide is useful for measurement practice, but its nine-second/ten-minute worked duty cycle is arithmetically inconsistent (the duty cycle is 1.5%). Those calculated lifetime figures are not calibration targets. Only the consistent SF7 power values in Callebaut Table 7 are used; several other technology columns appear to contain scale errors.

Averaging can fail for small batteries or long reports. One 180 mA transmission lasting one hour consumes 180 mAh. A battery with only 100 mAh available cannot finish that transmission, even if the daily average suggests 13.33 hours of life. The calculator blocks that configuration. The actual depletion time depends on whether discharge begins in sleep, wake, sensing, connection or transmit. For lifetimes covering fewer than ten reporting cycles, resolve discharge through the actual sequence of states and include self-discharge during each state. The plotted discharge and animation remain cycle averages, so they cannot predict the first brownout within a report. The ten-cycle warning threshold is a modeling heuristic; demands for tighter accuracy can require many more cycles.

The interval chart leaves intentional gaps at exactly zero load, where this charge-depletion model has no finite lifetime. Loaded points are finite, while calendar life and voltage cut-off still limit practical service.