KTP-762 Implemented Linear Quantile Regression Model

[egordm]

Following changes/features were added:

• New LinearQuantile model. Used by specifying model="linear_quantile" in PredictionJob
• Added feature filtering to ignore features deemed irrelevant.
• Added data imputer transform to replace nans in data as linear model can not work with these.
• Added tests for Linear Quantile Model.

Notes:

• The model training quite a bit slower than XGBoost Quantile.
• For each quantile a new SKLearn QuantileRegression model is trained.
• Adding more features significantly impacts training time.

Here are some insights from the comparison notebook:

Feature Importance

Feature importance is more balanced now, possibly also because of the feature filtering:

Model also sees that AMI, AZI etc features make a nice base signal.

Forecasting

• Model follows the load target more closely.
• The quantiles do not cross.
• The biggest quantile (0.9 in this image is better at forecasting peaks.
• There are more false positives.
  

[FrankKr]

Looks good Egor!
Two minor comments.
And if you are feeling clean-code-minded, you could solve the issues SonarCloud lists:
https://sonarcloud.io/project/issues?id=OpenSTEF_openstef&pullRequest=529&resolved=false&inNewCodePeriod=true

[egordm]

Reevaluated benchmark for the linear model after recent change (code style fixes). And change in y scaling:

[sonarcloud]

Quality Gate passed

Issues
3 New issues
0 Accepted issues

Measures
0 Security Hotspots
96.7% Coverage on New Code
0.0% Duplication on New Code

See analysis details on SonarCloud