Python-based package for performing prediction tasks based on numerical weather forecasts.
The package is based on the assumption that there is some numerical target you want to predict based on NWPs and is structured around simplifying this problem, including flexible support for different machine learning models, hyper parameter optimization and cross validation. It is assumed that the target is spatial (i.e. has a geographic position). The intended use is for wind power prediction but could be used to predict any variable based on spatial data (e.g. ice cream consumption or solar power production)
The target variable is defined by its geographical position and measurements. Location information should be a in a CSV file with three columns, the location id, latitude and longitude. Measurments should be a csv file where the first column is the time, and the remaining columns are production values for the locations over time.
Data is downloaded based on the location csv's. The coordinate file determines which coordinates are downloaded while the measurements CSV file determines what date intervals will be downloaded.
All configuration is based on python files. There are three main configurations:
- Dataset layout: This determines who the forecasts are converted into dataset, how they are windowed, what measurement to predict and so on
- Variable definition: This defines what variables from the NWP should be used and how they should be encoded.
- Training config: This defines how the training should be performed, such as cross validation strategy
First create a coordinate specification csv (assuming a POSIX system)
This creates a csv with two rows, the first is the header with column names and the second is the coordinates
Now use this coordinates file to download the NWP data: