Implementation of the paper Simulation-assisted machine learning for operational digital twins. In this paper we propose a method to make predictions overcoming the following problems:
- Not enough data exist to train machine learning models
- No data exist to calibrate process-based models in a new location
- Data exist but they are on different resolution than what machine learning and process-based models expect
We evaluate this method with a case study of pasture nitrogen response rate prediction in New Zealand. The evaluation includes different microclimates, scenarios of sampled/unsampled locations, and different amounts of data included in training. For the prediction models we used random forest and tuned it with bayesian optimization.
Before running the code, change the paths defined in the main function inside run.py to point to the simulated/weather/nitrogen response rate data, and also where to save the results.
Given the following example paths:
scenario_name: NZ_Cloverpreprocessing_results_path: preprocessing_results_pathml_results_path: ml_results
the following directory hierarchies will be created:
----preprocessing_results_path/NZ_Clover
|
----global
| |
| ----known
| | |
| | ----train
| | |
| | ----Clim1.csv
| | |
| | ----Clim2.csv
| | |
| | ---- ....
| |
| |----unknown
| |
| ----train
| |
| ----Clim1.csv
| |
| ----Clim2.csv
| |
| ---- ....
|
----regional
| |
| ----known
| | |
| | ----train
| | |
| | ----Clim1.csv
| | |
| | ----Clim2.csv
| | |
| | ---- ....
| |
| |----unknown
| |
| ----train
| |
| ----Clim1.csv
| |
| ----Clim2.csv
| |
| ---- ....
|
----local
| |
| ----known
| | |
| | ----train
| | |
| | ----Clim1.csv
| | |
| | ----Clim2.csv
| | |
| | ---- ....
| |
| |----unknown
| |
| ----train
| |
| ----Clim1.csv
| |
| ----Clim2.csv
| |
| ---- ....
|
----location
|
----test
|
----Clim1.csv
|
----Clim2.csv
|
---- ....
----ml_results
|
----NZ_Clover.zip