Twisp area is not reporting a Day 2 forecast

[vonw]

From: Friedman, Beth (ECY) <[email protected]>
Sent: Thursday, January 20, 2022 8:01 AM
To: Thorpe, Farren (ECY) <[email protected]>
Subject: one more forecast issue

Hi Farren,

One more minor forecast issue that I noticed—the Twisp area (AQS code 530470016) isn’t reporting a Day 2 forecast and based on the in the WSU ML forecast file it’s not included—is there anything in the WSU code pointing to a file we can update on our end with site locations or…?

-Beth

[vonw]

Email trail after initial issue

Friedman, Beth (ECY) [email protected]
Fri 1/21/2022 12:31 PM

To: Walden, Von P.
Cc: Thorpe, Farren (ECY) [email protected]; Fernandez Valdes, Ana Carla
Thanks Von and Ana Carla. There is a monitor in Twisp—I’m wondering if some of the confusion is due to the Twisp monitoring relocating in fall 2020 to a new location approximately 1 mile south of the original location. I’m guessing that previous Twisp data from that monitoring site (AQS ID for the previous site located on Glover St: 530470009, it’s been active since at least 2007) was used to train the ML model—and that data should be sufficient to apply to the forecast for the relocated site. I can also provide that data if you need it, or it is also in AQS.

Thanks for looking into this!

Beth

From: Walden, Von P. [email protected]
Sent: Friday, January 21, 2022 11:35 AM
To: Friedman, Beth (ECY) [email protected]
Cc: Thorpe, Farren (ECY) [email protected]; Fernandez Valdes, Ana Carla <[email protected]>
Subject: FW: one more forecast issue

THIS EMAIL ORIGINATED FROM OUTSIDE THE WASHINGTON STATE EMAIL SYSTEM - Take caution not to open attachments or links unless you know the sender AND were expecting the attachment or the link

Beth,

Please see the email below from Ana Carla. The ML model must have a good time series of past data from which to base its forecasts on. Do you know if there is a monitor in Twisp that might not be part of the AQS? If so, how can we get access to the data, so that we can train the ML model for this location?

Von

From: Fernandez Valdes, Ana Carla [email protected]
Date: Thursday, January 20, 2022 at 3:52 PM
To: Walden, Von P. <[email protected]>
Subject: Re: one more forecast issue

Hello Von,
Since I wasn't involved in the early stages of this project, I checked among all the outputs, and that specific site has never been forecasted in the ML model. I went further to check why it wasn't included, I first checked the AIRPACT data (the place where the ML model gets the majority of site input data from), and what I found was that the Twisp (530470016) site is not included in any of the datasets. I did a quick search on the AQS EPA website, and what I found was that the Twisp region had a PM2.5 monitor back in 2004-2005 (it had a different AQSID) but it is no longer active (see attachment). I also found this Air Monitoring Data Quality Assessment document from WA Department of Ecology, where that site is mentioned as “FRM-like” PM2.5 data (page 34). In that section, it specifically states that "PM2.5 data collected using nephelometers cannot be used for NAAQS compliance determinations". Therefore, I suspect that's the reason why that specific site has never been included in the ML model.

I hope that answers the question,
Best,
Ana Carla

Ana Carla Fernandez Valdes
Research Assistant
Laboratory for Atmospheric Research
Washington State University
From: Walden, Von P. <[email protected]>
Sent: Thursday, January 20, 2022 1:56 PM
To: Thorpe, Farren (ECY) <[email protected]>; Friedman, Beth (ECY) <[email protected]>
Cc: Fernandez Valdes, Ana Carla <[email protected]>
Subject: Re: one more forecast issue

Beth (and Farren),

I’ve forwarded this along to Ana Carla. We’ll be back in touch soon with an answer.

Von

[vonw]

Response from Ana Carla

Fernandez Valdes, Ana Carla
Thu 2/3/2022 12:19 PM

To: Walden, Von P.
Thank you so much for your feedback.

On another note, I think I know understand what is happening with the Twisp site.
You were right about the missing data for 2020. On the 2020 AIRPACT's online repository the "840530470016" site is missing but when you look at the 2021's data that site is there.
Therefore, what happened is since the latest trained model was based on 2020's data, when the forecasting for 2021 happens, the prediction model is trying to match the sites from the trained model with the new data, and any site outside of the scope of the 2020 list is skipped.
The solution to that problem is to re-train the model with 2021 data (something that is due soon), but I will have to discuss it with Kai first due to some technical issues that I am not sure how are going to work (I can explain those later during a meeting).

Best,
Ana Carla

Ana Carla Fernandez Valdes
Research Assistant
Laboratory for Atmospheric Research
Washington State University