AIR PPM Hackathon – Challenge Description

The minimum prize pool is USD 3,000 (three thousand US dollars).

Training data for model development is available at:
https://drive.google.com/drive/folders/1vZBjKSeMab-f8owLfjQyEEwaPTYV-hRt?usp=sharing
You must preprocess those raw files into the JSON schema described below before using them.

1. Reference Implementation

A template project is available at:
https://github.com/AIRPPM/submission_example

In the main script, find the line starting with TODO and replace the random-value generator with your PM₁₀ forecasting model.

2. Input & Output Formats

Input JSON (data.json)

The input must be an object containing a list of cases.
Each case includes:

• case_id: unique identifier for this scenario
  
  

• stations: list of measurement stations, each with:
  
  

  • station_code: alphanumeric code
    
    

  • longitude, latitude: geographic coordinates
    
    

  • history: past hourly readings of PM₁₀:
    
    

    • timestamp: when the measurement was taken
      
      

    • pm10: recorded PM₁₀ value
      
      

• target: location and time to start the forecast
  
  

  • longitude, latitude
    
    

  • prediction_start_time: beginning of forecast period
    
    

• weather (optional): array of METAR-style observations to enrich your model
  
  

Example:

{

  "cases": [

    {

      "case_id": "case_0001",

      "stations": [

        {

          "station_code": "StationA",

          "longitude": -74.0060,

          "latitude": 40.7128,

          "history": [

            { "timestamp": "2025-01-01T00:00:00", "pm10": 42.5 },

            { "timestamp": "2025-01-01T01:00:00", "pm10": 38.1 }

          ]

        }

      ],

      "target": {

        "longitude": -74.0060,

        "latitude": 40.7128,

        "prediction_start_time": "2025-01-03T00:00:00"

      },

      "weather": [

        { "date": "2025-01-01T00:00:00", "tmp": "+0050,1", "wnd": "260,1,N,0030,1" }

      ]

    }

  ]

}

In this example, one case is defined for a station in New York City.
The history array contains two hourly PM₁₀ readings, and the forecast will start at midnight on January 3, 2025.

Output JSON (output.json)

The output must be an object with a list of predictions.
Each prediction links back to the input case and provides 24 hourly PM₁₀ forecasts.

Example:

{

  "predictions": [

    {

      "case_id": "case_0001",

      "forecast": [

        { "timestamp": "2025-01-03T00:00Z", "pm10_pred": 48.7 },

        { "timestamp": "2025-01-03T01:00Z", "pm10_pred": 52.3 }

      ]

    }

  ]

}

Here, the output for case_0001 contains two example forecast points.
Your model must produce 24 entries per case; fewer entries will trigger an error.

3. Docker Packaging

Build a Docker image named, for example, air-pm10-forecast.
It must contain your code and dependencies so that it can be run with the commands below.

Build:

docker build -t air-pm10-forecast .

Run example:

docker run --rm \

  -v /local/data/data.json:/data/data.json \

  -v /local/data/landuse.pbf:/data/landuse.pbf \

  -v /local/data/output.json:/data/output.json \

  air-pm10-forecast \

  --data-file /data/data.json \

  --landuse-pbf /data/landuse.pbf \

  --output-file /data/output.json

The container must exit with status zero only if all cases produce exactly 24 hourly forecasts;
otherwise it must exit with non-zero to indicate failure.

4. Submission Guidelines

Upload a ZIP archive containing:

• A Dockerfile at the archive root, which specifies how to build your image.
  
  

• All source code and resources required to reproduce that build context.

• Link to submission form:
  https://docs.google.com/forms/d/e/1FAIpQLSfsO57MN36h9Qr7ZZY6XGC8e7G2iG_nuCblUEzUZMZNJLka4Q/viewform?usp=dialog
  
  

When unzipped, running docker build -t your-image-name . in that directory must succeed without additional modifications.

5. Evaluation

Submissions will be evaluated on Mean Squared Error (MSE) against a hidden test set.
Lower score wins; ties are broken by earlier submission time.