How can I make a battery-powered ESP32 Wi-Fi sensor last for months instead of days?

Battery life is capacity divided by average current, so work out the average. An ESP32 with the radio active draws on the order of 100 mA, with transmit peaks of a few hundred mA according to the datasheet; 2000 mAh at that level lasts about a day, which is what you see. In deep sleep the chip itself needs around 10 µA with only the RTC timer running.

Now duty-cycle it. Suppose each wake-up takes 3 seconds at an average of 100 mA: that is 300 mA·s, or 0.083 mAh. With 144 wake-ups per day the active part costs 12 mAh per day, and sleep adds 0.01 mA × 24 h = 0.24 mAh. At about 12.2 mAh per day, 2000 mAh lasts roughly 160 days before allowing for self-discharge and usable capacity.

The calculation shows where the energy goes: almost all of it is the awake time, so shortening each wake-up matters more than anything else. A plain connect with DHCP often takes 2 to 5 seconds. Typical savings: use a static IP address to skip DHCP; store the access point's channel and BSSID in RTC memory and pass them to WiFi.begin() so the scan across all channels is skipped; read the sensor before turning the radio on; and keep the exchange short, since a TLS handshake alone can add a second or more.

Getting the awake time from 3 seconds to 1 second roughly triples the battery life in the example above.

The board often matters more than the chip. The 10 µA figure is for the bare chip or module. A typical development board adds a USB-to-serial chip, a power LED and a linear regulator with a quiescent current of several mA, so it draws milliamps while "asleep". At 5 mA the sleep current alone uses 120 mAh per day, and the battery is empty in about 16 days however clever the firmware is.

For a real battery node, use a bare module or a board designed for low power, with a regulator whose quiescent current is in the microamp range, no always-on LED, and sensors that are powered down during sleep. Measure the sleep current with a meter on the µA range before optimising anything else.

Watch the voltage as well as the capacity. The module is specified for about 3.0 to 3.6 V and a lithium cell goes from 4.2 V down to about 3.0 V, so a regulator is required, and a common linear type with around 1 V of dropout cannot hold 3.3 V from a single cell at all. Choose a low-dropout regulator rated for the transmit peaks (at least 500 mA) with low quiescent current, or a buck-boost converter, and put a bulk capacitor close to the module so the peaks do not cause brownout resets on a partly discharged cell.

Use a protected cell or a protection circuit, charge it only with a proper lithium charger, and do not discharge below the limit in the cell's datasheet.

If the node must stay reachable instead of waking on a timer, deep sleep is not an option, but the Wi-Fi power-save modes are. In modem sleep the station keeps its association and turns the radio on only for beacons at the DTIM interval, bringing the average down to a few tens of mA; automatic light sleep reduces it further, at the cost of latency.

Those modes suit mains-backed or large-battery devices. They do not get anywhere near the duty-cycled deep sleep figures, so for a sensor that reports every 10 minutes, timer wake-up from deep sleep is the right design.