Apollo Irradiance Forecasting System¶
Documentation¶
Overview: The What and Why of Apollo¶
Apollo is an open source system for machine learning from numerical weather forecasts. Its purpose is to facilitate data collection, model training and evaluation, and the serving of predictions.
System Architecture¶
System Architecture of Apollo
Data Ingestion Subsystem¶
Todo
TODO
Modeling Subsystem¶
Todo
TODO
Storage Subsystem¶
Todo
TODO
Service Subsystem¶
Todo
TODO
Quickstart: Running Apollo with Docker¶
git clone https://github.com/cbarrick/apollo.git
docker build apollo -t apollo
docker run -it apollo
Installing Apollo Locally¶
Python API Reference¶
Modeling Framework¶
Apollo’s core API is the modeling framework.
Model Classes
apollo.models.Model |
|
apollo.models.IrradianceModel |
|
apollo.models.NamModel |
Utility Functions
apollo.models.list_models |
|
apollo.models.list_templates |
|
apollo.models.load_model |
|
apollo.models.load_model_at |
|
apollo.models.load_model_from |
|
apollo.models.make_estimator |
|
apollo.models.make_model |
|
apollo.models.make_model_from |
NAM Forecast Data¶
Apollo uses the North American Mesoscale (NAM) forecast system, a numerical weather simulation produced by the National Oceanic and Atmospheric Administration (NOAA). Apollo can be configured to collect NAM forecasts as training data for machine learning models.
Data Access
apollo.nam.download |
|
apollo.nam.open |
|
apollo.nam.open_range |
|
apollo.nam.CacheMiss |
Geographic Coordinates
apollo.nam.NAM218 |
|
apollo.nam.proj_coords |
|
apollo.nam.slice_geo |
Useful Constants
apollo.nam.ATHENS_LATLON |
|
apollo.nam.PLANAR_FEATURES |
See also
- NAM Home Page
- Detailed documentation of NAM.
- National Centers for Environmental Information
- Access to raw NAM forecast data.
- Inventory of File nam.t00z.awphys00.tm00.grib2
- Catalog of variables included in NAM forecasts. (Apollo does not support every variable.)
Feature Extraction¶
apollo.time_of_day |
|
apollo.time_of_year |
|
apollo.is_daylight |
Metrics¶
apollo.metrics.all |
|
apollo.metrics.mae |
|
apollo.metrics.r2 |
|
apollo.metrics.rmse |
|
apollo.metrics.stdae |
Visualizations¶
Apollo includes several visualization routines.
apollo.date_heatmap |
|
apollo.date_heatmap_figure |
|
apollo.nam_figure |
Data Access¶
Apollo stores models and datasets in the Apollo database. The database is a regular directory specified by the APOLLO_DATA environment variable, defaulting to /var/lib/apollo. In the Apollo Docker image, the database is a volume mounted to /apollo-data.
apollo.path |
REST API Reference¶
Todo
Document
Command Line Reference¶
Apollo provides a command line toolbox for managing forecast data, developing forecast models, and administering the server. All Apollo commands sport a --help option with detailed descriptions. This documentation provides an overview of the commands.
apollo¶
Summary¶
The Apollo CLI toolbox.
Usage¶
apollo [-h] [--quiet | --debug | --log LEVEL] COMMAND ...
Description¶
The apollo command provides a toolbox style CLI, like git. The root apollo command takes a single required argument, COMMAND, giving the subcommand to execute. Optional arguments that come before the subcommand are applicable to all subcommands, while arguments that come after are specific to the subcommand.
apollo ls¶
Summary¶
List items within the Apollo database.
Usage¶
apollo ls [-h] [COMPONENT]
Description¶
The apollo ls command is for listing different items stored in the Apollo database. You can optionally specify a component to list only those items.
Components include:
models: The trained models.templates: Templates for training new models.nam: Available NAM forecasts.
Examples¶
List everything in the database:
$ apollo ls
models/linear-nam-uga
models/xgboost-nam-uga
...
templates/linear-nam
templates/xgboost-nam
...
nam/2017-01-01T00Z
nam/2017-01-01T06Z
nam/2017-01-01T12Z
nam/2017-01-01T18Z
nam/2017-01-02T00Z
nam/2017-01-02T06Z
nam/2017-01-02T12Z
nam/2017-01-02T18Z
...
List only NAM forecasts:
$ apollo ls nam
2017-01-01T00Z
2017-01-01T06Z
2017-01-01T12Z
2017-01-01T18Z
2017-01-02T00Z
2017-01-02T06Z
2017-01-02T12Z
2017-01-02T18Z
...
Thesis¶
Note
These are my thesis chapters and won’t appear in the final docs.
Introduction¶
Todo
TODO
System Design¶
Use Cases¶
Todo
Copy from Overview: The What and Why of Apollo.
System Architecture¶
Todo
Copy from Overview: The What and Why of Apollo.
The NAM-UGA Dataset¶
NAM¶
Todo
- What is the NAM forecast model?
- NOAA
- GRIB
WRF-NMM¶
Todo
NAM is an instance of the Weather Research and Forecasting (WRF) nonhydrostatic mesoscale model (NMM)
NAM-UGA¶
Todo
- Feature subset
- Geographic subset
- Data collection period
- netCDF
Experiments¶
Description of Train and Test Data¶
Todo
- Simple test-train split.
- Train data from 2017.
- Test data from 2018.
- Calendar heatmap graphic.
Evaluation Metrics¶
Todo
- MAE vs. MSE
- MAPE vs. sMAPE
- R^2
- day-night vs day-only
Preprocess & Feature Selection¶
Todo
- What features are useful?
- Computed time-of-day and time-of-year features.
- Training with day-night vs day-only.
Empirical Learning Curve¶
Todo
- Error vs amount of training data used.
- Linear regression vs. Random Forest vs. GBTs.
Reference Time vs Forecast Time¶
Todo
- Error for each (reftime, forecast time) pair.
- Heatmap.
Conclusions¶
Todo
TODO
Acknowledgements¶
Apollo was developed at the UGA Institute for Artificial Intelligence in Athens, GA and funded by the Georgia Power Company.
The primary authors of Apollo are:
- Chris Barrick
- Zach Jones
- Aashish Yadavally
- Dr. Fred Maier
A more complete listing of code contributors can be found here.
The research and development of Apollo was directed by:
- Dr. Fred Maier (UGA)
- Dr. Khaled Rasheed (UGA)
- Will Hobbs (GA Power)
Apollo’s logo was designed by Freepik and is used under the terms of the Flaticon basic license. The logo was obtained from Flaticon.
The Apollo wordmark was added to the logo and uses the Fira Sans font. Fira Sans was designed by bBox Type and is used under the terms of the SIL Open Font License (OFL).