Artificial Intelligence for Weather Forecasting in Comparison With Current NWP Models
Advisor: Anmol Bhardwaj
Weather forecasting is essential to disaster preparedness, aviation safety, agriculture, and long-term climate planning. The traditional Numerical Weather Prediction systems (NWP) rely on physical equations and formulas. Recently, artificial intelligence-based forecasting models have emerged, using large historical data sets and inductive reasoning instead. This project investigates the strengths and limitations of AI-based weather models compared to traditional NWP systems. By comparing forecast accuracy across different variables such as temperature, precipitation, and extremity, I investigate where the AI models outperform and underperform the NWP systems — notably long-range stability and in rare events. I will also examine how the strengths of both can be combined and best integrated in hybrid models.
Project Blog
Post 1: Introduction: To What Extent Can Man Predict Nature?
This week, I have been working on my Senior Project, which explores the intersection of Artificial Intelligence and Weather Forecasting. I was drawn to this project because I have always been fascinated with meteorology and how it saves lives. Being someone prone to worry, even more so in my younger years, meant that I often stayed up worrying about tornadoes. I remember my absolute horror reading a children's book set in the 80s describing a typical encounter with a tornado when it was explained that the warning time for the tornado was THREE MINUTES. Immediately after, I remember the wave of relief upon learning that in the modern day (2015 at the time) the warning time was 15 minutes. It still wasn’t much, but according to my calculations as a 7 year old, it was enough time to run to the basement with cherished belongings and people. Earlier warnings have been made possible bythe advances of our understanding of meteorology and meteorological technology. They promise to become even better with more advances in technology, such as the introduction of AI to forecasting. That is what my project is about.
A map of deadly tornadoes in the state of Virginia from 1950 - 2025, The NewsLeader
I hope that through my research into AI and weather forecasting, I can better understand the emerging technology of meteorology predictions and forecasting. I hope that I can give myself some assurance that we will have even warning before the very land we live on spontaneously crumbles or gets buried under floodwaters. I hopeemerging technology will help us understand the natural world better to make a safer world for the next generations.
The last deadly tornado in Northern Virginia, the region I spent 96.2% of my life in, was in 1979, 29 years before I was born. In the last decade, there have been 5fatalities from tornadoes from 2 deadly tornadoes in the State of Virginia (The NewsLeader). I hope that eventually, there will be zero fatalities for tornadoes, and that what began as a general anxiety about the unpredictability of the natural world can evolve into an informed opinion on technology and weather.
(Evolution Of A Tornado)
Bibliography:
“A History of Twisters: Tornadoes in Virginia Since 1950.” Virginia Tornado Archive, Data Central, The NewsLeader, https://data.newsleader.com/tornado-archive/virginia/deadliest/
“Evolution of a Tornado.” Wikimedia Commons, uploaded by User:DBasner, 11 Nov. 2007, https://upload.wikimedia.org/wikipedia/commons/8/86/Evolution_of_a
_Tornado.jpg. Accessed 2 Mar. 2026.
“Five Ways NOAA Scientists Helped Improve Human and Marine Life in 2025.” NOAA Research, U.S. Department of Commerce, National Oceanic and Atmospheric Administration, 25 Feb. 2026, https://research.noaa.gov/five-ways-noaa-scientists-helped-improve-huma n-and-marine-life-in-2025/
. Accessed 2 Mar. 2026.
Post 2: Blog 1: Forecasting Innovation and Bayesian Interference
This week it felt warmer outside. The forecast is for (mostly) warm weather in the coming week, which makes me grateful to have a forecast that gives me something to look forward to.
In recent forecast technology news, the Weather Research and Forecasting Innovation Reauthorization Act of 2026 was approved by the Senate Commerce Committee on March 4, 2026. This bill aims at updating NOAA's weather forecasting capabilities, modernizing radar/satellite systems, and improving disaster warnings. One of the primary actions this legislation would take is integrating more artificial intelligence into existing weather systems. As a member of the weather-obsessed public, this legislation brings me joy (and also concern because why was improving forecast technology even up for debate, especially in places with extreme weather events (ie California)).
I have continued my background research. I learned that using something called Bayesian interference to understand weather conditions could be a good way to handle uncertainty (Jian et al 2021). Bayesian interference updates the probability of a hypothesis as more information is made accessible. It begins with a set belief (often called a “prior distribution”). It considers new probabilities as they are added and calculates an updated probability (Taboga 2021). Bayesian interference implements something called Bayes' Rule which I still do not completely understand, but that’s okay. I’m still working on it and will get you all an update next week.
For now, enjoy the diagram from Wikipedia I consulted in an attempt to understand Bayes' Rule.
By Cmglee - Own work, CC BY-SA 4.0, https://commons.wikimedia.org/w/index.php?curid=143581610
Sun Jian, Cao Zhuo, Li Heng, et al. “Application of artificial intelligence technology to numerical weather prediction.” J Appl Meteor Sci, 2021, 32(1): 1-11. DOI: 10.11898/1001-7313.20210101.
Taboga, Marco (2021). "Bayesian inference", Lectures on probability theory and mathematical statistics. Kindle Direct Publishing. Online appendix. https://www.statlect.com/fundamentals-of-statistics/Bayesian-inference.
Week 2: A Dig For Data and The Relevance of Marine Weather
This week, I focused on refining my method in my senior project. I have been looking into exactly what specific available models I should use to draw my conclusions from. In addition, I have been searching for large historical data sets I can input into models to test my hypotheses. After some searching, I discovered the NOAA has a fairly large and detailed set of historical weather data. A two month log of weather in my zip code alone requires a bulky, compressed file. Due to an issue with their servers, I was unable to properly download said bulky, compressed zip file, however, I do now know where to acquire a historical data set for my testing use. Additionally, the NOAA website includes specific historical data pertaining to severe weather and storm events.
Interestingly, the NOAA also has data on marine weather. Something I hadn't considered before is how marine weather patterns could have an affect on land weather patterns, even more so for Northeast Virginia, as we are relatively close to marine environments. Marine weather can create fog, trigger thunderstorms, and cause severe rainfall. The NOAA has an additional log for marine data specifically, and I will be taking that data into account as well when I test my models. The marine data might affect Fairfax County less because it is not a coastal county (although we do border one); if that happens, I will report on it if necessary.
Works Cited:
National Centers for Environmental Information. Climate Data Online: Data Tools. National Oceanic and Atmospheric Administration, https://www.ncei.noaa.gov/cdo-web/datatools. Accessed 12 March 2026.
Week 3: Technical Difficulties and Refined Methodology
This week I focused on setting up my methodology, which I have recently refined to make more actionable.
First, historical weather data will be collected from NOAA, including variables such as temperature, wind speed, pressure, and precipitation. Next, historical forecasts from the Global Forecast System (GFS) will be obtained. I will check the accuracy of the historical forecasts to determine the functionality. An artificial intelligence model will then be implemented using a cloud-based platform such as (GraphCast). This study will reference advanced AI models such as GraphCast to provide context for modern forecasting techniques.
The location used will be Herndon, Virginia. Forecasts will be evaluated at different time intervals, and accuracy will be measured. Finally, results will be analyzed to determine whether AI models can match or improve upon traditional NWP forecasts, especially at longer prediction times.
In terms of implementing my method, I have taken two steps. First, I went onto the National Oceanic and Atmospheric Administration (NOAA) website in order to find the GFS (Global Forecasting System) historical forecasts. Historical forecasts, in this context, meaning historical analysis and predictions of historical data. Second, I have been setting up GraphCast, which is the most suitable model for my needs that can safely function from a home computer. GraphCast does not require a supercomputer like the models weather stations use (though it does take a very long time to load, simply because it is dealing with so much data). I have been troubleshooting errors with GraphCast which are as plentiful as they are annoying. I have managed to get it running. Now, I simply have to make it run in a non-erroneous manner.
Overall, this week has been less about reviewing literature and more focused on setting up the technical groundwork for my method.
Week 4: First Draft Due Date
This week, I have been working towards finishing up my first draft: Intro, Literature Review, and Method. This has consumed me all week until about 5 minutes ago from when I wrote this. Specifically, I expanded my literature review. I looked into the general criticisms of the current U.S. Operational Numerical Weather Prediction, and suggestions for improvement. I reviewed an article that, instead of explaining the drawbacks of Numerical Weather Prediction systems in general, specifically criticized that of the United States. The article mentioned that the systems in the United States were behind that of other places, such as the n European Centre for Medium-Range Weather Forecast. The article gave recommendations on strengthening the NWP systems in a less artificial-intelligence-focused manner. The article also gave me more motivation to look into how other NWP models would work, giving me a less America-centric view.
This is one of many articles and papers I have read this week, although it was one that particularly caught my eye (although it did still contain a sizable amount of AI implementation suggestions). Overall, I've expanded my literature to a certain extent. Additionally, I looked into how my research would affect other fields, such as agriculture and aviation. Did you know that about 70–80% of year-to-year crop yield variability is influenced by weather? In addition, did you know that a single 1% improvement in weather forecast accuracy can save the aviation industry hundreds of millions of dollars per year? These facts emphasize the value of using AI to improve weather forecasting.
Week 5: Challenges
This week I have been working on expanding my research. I have run into some issues. For one, the GraphCast model I was examining has crashed and has stopped working. I am sure it is just a temporary bug; I am working on getting it up and running again. It did work before, so I am not completely void of hope. I still want to try and expand my literature review as I feel that it is not enough. I have come to a point in my literature review that I am struggling to find more (diverse) sources online free of cost. I want to diversify my sources, so I think I will try to take to the local library and/or a bookstore and see if they have physical sources I can comb. I skimmed some books at my local bookstore this week, so I am hopeful of the output that should provide. If I still feel that my sources are either low in quantity or homogenous, I might go to a local university library and see what they have in their collection. Admittedly, I have hit a bit of a roadblock in my project, one that I intend to be temporary. I will continue to try to overcome this.
Week 6: Attempts to Find New Literature
Fun Fact: Did you know that in 2023, GraphCast predicted Hurricane Lee three days before the traditional forecasting models did?
This week was also a little slow in terms of my research. I did read some more sources, including one where I learned more about the systems of the past and how they compare. I learned, specifically, about analog forecasting, the system used before computers. Forecasters would search for the closest match between the weather map of that day and a weather map from the past (an analog). Neural network forecasting (or some at the very least) is based on the principles of analog forecasting.
In addition, I also investigated the cost discrepancies between model types. I learned that GraphCast is about 1000 times cheaper in terms of energy efficiency than traditional models. I already understood that artificial intelligence models were cheaper in the long run, but require more initial costs to set up. I learned that AI-driven tools can generate results from a laptop, in contrast to the multi-million dollar supercomputers required to solve difficult equations and crunch large numbers. I read that the machine learning method developed by Oxford has proven more effective than other methods of rainfall prediction.
I am going to see if I can find more in-depth cost breakdowns of the models to use in my research. So far, cost metrics are usually a subsection within articles whose main focus is elsewhere. Hence, I am going to attempt to find a source where the main subject is the cost breakdowns and discrepancies between AI and NWP models (and maybe the ancient analog system, if relevant).
Sources:
“Neural Net Forecasting Goes Old-School.” Bulletin of the American Meteorological Society, vol. 101, no. 4, 2020, p. 284. JSTOR, https://www.jstor.org/stable/27028166. Accessed 17 Apr. 2026.
Donback, Natalie. “Traditional Weather Forecasting Is Slow and Expensive. Ai Could Help.” Grist, 3 Jan. 2025, grist.org/extreme-weather/how-ai-could-help-predict-climate-fueled-extreme-weather/. Accessed 17 Apr. 2026.
“AI Can Now Outperform Conventional Weather Forecasting – in under a Minute, Too.” World Economic Forum, 14 Dec. 2023, www.weforum.org/stories/2023/12/ai-weather-forecasting-climate-crisis/?utm_source=chatgpt.com. Accessed 17 Apr. 2026.
Week 7: The World is Chaotic
This week I have been largely working towards my final product draft 1. I explored more about the specifics regarding the different uses and applications of weather forecasting in various fields. I learned of the specifics of weather forecasting’s importance to agriculture. Anticipation of heatwaves allows farmers to change their irrigation schedules accordingly in order to maintain sufficient soil moisture and prevent heat stress in crops; anticipation of heavy rainfall allows farmers to delay plantation or harvest, intending to avoid seed washout, soil compaction, and spoilage of mature product. Wind predictions are also apparently useful for deciding when to spray pesticide and herbicide, with intention to lessen drift and optimize effectiveness. I also learned of the military use of meteorology; a subfield called military meteorology. Military meteorology is exactly what it sounds like – attempting to use the chaos of the natural world to one's advantage. I never really considered that weather forecasting could be used for offensive action, but now I know.
In addition, I understood the butterfly effect. I knew before that it was when a small, seemingly insignificant event causes a much larger consequence as a result of little mishaps accumulating over time. In the context of my research, it would be in regards to how the AI-based and ML-based models struggle with long-term weather prediction because little relatively insignificant inaccuracies accumulate over a large time frame to provide us with large and very much significant inaccuracies. However, I now learned that the butterfly effect is the idea that a a distant butterfly flapping its wings creates a miniscule instability in the atmosphere, a miniscule instability which becomes a much larger instability. (side note, chaos theory is really interesting, and it’s great I have an excuse to use it.) The goal of the theory was to explain that weather can not be truly and wholly forecasted, due to the inherent chaos of the atmosphere. I suppose what forecast technology we have now and are trying to develop simply tries its best to forecast.
Week 10: Final Edits
I have been working on editing my final draft. Overall, positive feedback, marginal edits required however. As I might have mentioned earlier, due to the limitations of my main method, I decided to demote it to a secondary method a while back. Under advice, my new methodology drew upon established, published data and data evaluations to draw my own conclusions. I also checked out the data and evaluations from case studies of certain weather events. Running the GraphCast model proved difficult, so under advice, I tested another open source model. This one is an LSTM model - a long short term memory model.
Traditional RNNs struggle with remembering information over long periods of time (just like me); LSTMs are specialized to remember context. It seems like I made my paper more confusing. I fixed it by explaining the other methodology.
This blog was fun. I learned a lot.
Signing off,
Neya