CL System for Pressure calibration explained



On October 12th we quietly launched our new Continuous Learning (CL) auto-calibration system for Pressure. The system performs a quality control inspection on the raw station pressure sensor data from your AIR, and applies calibration corrections if needed. Here’s how it works:

The Pressure CL system begins by comparing your AIR’s pressure data to numerous trusted data sources in your immediate area. The data sources are used to define a reference sea level pressure value. The value is then adjusted to an equivalent station pressure using your station’s elevation and AIR’s height above ground level. As a final step, the system will analyze and QC the station pressure, and once confirmed, a calibration is computed and applied to your AIR. Subsequent station pressure observations from your AIR will be governed by this calibration. The system is run every day keeping your pressure data in perfect calibration forever.

Note: if you have made adjustments to your elevation to manually “calibrate” your pressure sensor, please feel free to set it back to the correct value. Doing so will result in odd pressure values at first, but give the CL Pressure system a day or so to run again, and it will fall into line.

If after a day or two your Pressure calibration still seems off, please let us know directly and we’ll dig into your data.

A few notes on Station Pressure versus Sea Level Pressure….

Sea Level Pressure = Barometric Pressure = Relative Pressure = air pressure adjusted to sea level

Station Pressure = Atmospheric Pressure = Absolute Pressure = air pressure measured by the sensor

There are two pressure values reported in the Smart Weather app. One of them is called station pressure - it’s also called “absolute pressure”; or “atmospheric pressure”. This is the value actually measured by the sensor in your AIR. This value is not of much use to people, unless for some reason you want to know what the actual pressure is at your house (maybe for brewing beer or baking bread ?).

The other pressure value reported is “sea level pressure”; - it’s also called “relative pressure” or “barometric pressure”. This value is NOT what the sensor actually measures. Rather, it’s the value that would be measured if you were standing at mean sea level under the same sort of atmospheric conditions at your house. It’s adjusted or “normalized” based on the device’s total height above mean sea level. In the Smart Weather context, the “total height” is the sum of the “elevation” value that you set for your station location and the “height above ground” value that you set for your AIR device.

The value displayed in the main interface of the Smart Weather apps is sea level pressure. Why? Because that’s the most common way to report pressure and it’s the only way to facilitate comparison between locations with different altitudes. This lets you compare apples to apples. This is the value meteorologists care about. This is the value most weather services, websites and apps commonly report. This is also the value that most other home weather stations display. It’s also the value an analog mercury barometer should show.


REST API - "barometric_pressure" will be deprecated in favor of "station_pressure"
Air pressure is inaccurate (solved with CL system)
Continuous Learning - which variables?
Barometric Pressure way off (corrected once CL system is deployed)
pinned #2


So simple evan a Sea Hawks fan can understand.


I see.
My pressure is way better in comparison with official KNMI stations.

Great work WF. :+1:


great, but I still wonder what use this is. So you take measurements from other sensors around me, apply some sophisticated model, and calibrate my air unit so it matches the value of the model. Sounds great, but why do we then have a pressure sensor at all? you could simply apply the model without my sensor and show that data. (assuming the model updates fast enough to follow the pressure change).
Some sensors should only calibrated once, preferable against a standard, or perhaps once a year, not daily.


So the calibration ultimately adjusts the station pressure value to match what it really is at the AIR. I have made sure that my two collocated AIR units are at the same elevation in settings but they still show a difference of 0.003 inHg. Is this within the error of measurement of the sensor?


Hi @gizmoev . Glad to see your two stations now in cahoots. The spec on the pressure sensor in the AIR is accurate to within +/- 1millibar. As 1millibar = 0.02953 inHg …it’s safe to say that your difference of three one-thousands inch Hg (0.003 inHg) is well within the tolerance. In fact, if we weren’t so aggressive in showing three decimals points of inHg, you wouldn’t even know. Ha! :wink:


Hi @sunny Great questions!..and certainly ones that we anticipated before putting all the time and energy into building a smarter weather network, which is more than just a bunch of hardware. Your station feeds the network and the collective intelligence of the network can provide meaningful reciprocal value to each individual station. Together smarter, better.

Well, close but not quite correct. Data from your station is compared to the target values via least-squares polynomial regression, which does not match point-for-point but rather aims to bring the long-term averages in line with each other. Your AIR will continue to report unique pressure data from your exact location - the CL process ensures it’s the most accurate data possible.

That’s true, we could. And we may do that in the future for those unfortunate souls who don’t have a Smart Weather Station. But there is significant value in getting one-min data direct from your (calibrated) station in real time. And keep in mind, the ground truth data from your station can eventually improve forecast models, and the models plus the relevant comparison data informs your station keeping it in check.

Some sensors, maybe, but the sensors in weather stations definitely need periodic calibration. The best calibrations can drift due to many different reasons (age, conditions, physical movement, etc.). The sensors in your AIR & SKY are calibrated at the factory (this is the reference device we use for the pressure sensor:, which is NIST calibrated in a lab annually). But if you have the data and technology to perform calculated automated quality control more frequently, why wouldn’t you?


Great answers! thank you very much, it leaves just one burning question… if my sensors are calibrated so they better match the model values, doesn’t that disqualify them for being used as reliable input for the model?
My sensor might be perfect and the model slightly wrong, in which case it could be used as input for the model. If my perfect sensor is ‘calibrated’ to match the model, how is that being used to improve the model?


I image “numerous trusted data sources” is relatively easy for WeatherFlow in the US. How well does the system work for users in other parts of the world? I expect there is number of stations that have no trusted sources in the immediate area. What happens in this situation? It would be really good to know if CL calibration is being applied to a reading; is there any flags in the api data?


Correct, the more trusted reference points, the better the ability to define a mean calibration value. Keep in mind that we also evaluate multiple high resolution forecast models and have a comprehensive numerical weather prediction capability. At your location our CL system is currently using 8 different stations within 3km plus a handful of forecast models.

In areas where there are no other reference stations, the CL system is still better (than not having a system) as it leverages the hi-res forecast data to govern and QC the observations.

We agree! Working on enhancements to the app user experience to clearly indicate CL / calibration status for each sensor. No flags in the API yet…gotta crawl before we walk.


Sure, no problem. There might be some confusion here with the use of the term “model”. For clarity sake, let’s reserve “model” for reference to an individual “forecast model” such as the GFS or WRF. And use “CL system” to describe the collective processes we employ to both QC individual station data and eventually learn from the network.

The data sources for CL system calibrations are multiple and independent; specifically not informed by the data that they are being used to calibrate.

As the network grows with adequate density, the assimilation data can be used to validate or dispute the performance of numerical weather prediction models. In instances of dispute, applied machine learning can be used to understand data relationships and ultimately improve the underlying logic and calculations.


I am founding real data for actual pressure from my station very useful during the burning fireplace :slight_smile:

Once during previous year room was full of smoke because chimney return the smoke back. Usually everything work fine, but in that moment pressure was too low like never before. I found that fact later when I complain to chimney sweep. He suggested me always to look data about pressure (I am not using fireplace very often)…

So, if station pressure data from AIR will help me in not ruining romantic event, then this data will be very useful :slight_smile: Baked hot bread is great and brewing bear is excellent idea :slight_smile:


So if you want station pressure can you not just set your elevation to 0


Hi @coyyote . The solution is much easier…just look at ‘Station Pressure’ in the List View screen.

pinned #17


Hi there, my Air sensor’s pressure measurement

did something strange this morning. Pressure was pretty steady all night (clear skies, no synoptic weather systems moving in) and then suddenly the pressure increased by almost 3 millibars in only one minute, then remained steady as before (see attachment). Was this a calibration? There are two other weather sensors within a half a mile from me and neither of them saw this jump, therefore I believe it was my sensor. Staiton # 6079.

Am I correct in this assumption? Also, does this sort of thing happen often? My sensor has only been in operation for about 5 days, but this data jump is a little discouraging. Thanks!


It could also be that a calibration was applied to your station at that time.


Right, that’s what I figured, but my concern is how often are these applied/necessary? This is a pretty significant data jump and I’m hoping this doesn’t happen very often. Again, I’ve only been up and running for 5+ days.


It is consistently being check and will be applied as necessary. Once it settles down you should not notice any big changes.

I’m not sure why you are concerned that it is being calibrated. I would be concerned if it did not get calibrated.