Multi-step time-series forecasting of PV production using chained regressors from the scikit-learn Python library on the Liege Microgrid Open Data from Kaggle.
The aim of the project is demonstrating, among other things, the use of pipelines and regressor chaining for multi-step time-series forecasting. Regressor chaining allows implementing multi-output regression for all scikit-learn regressors that do not support it natively. Project also includes features engineering for time-series forecasting, data preprocessing, principal component analysis, hyper-parameters optimization with cross-validation, and other typical machine learning tasks.
Project is also concerned with testing of the stability of the newly introduced HalvingRandomSearchCV randomized search strategy for the optimization of model hyper-parameters. This new strategy is still experimental at the time of this writing. Comparison of the novel HalvingRandomSearchCV with the traditional RandomizedSearchCV is carried out, in terms of execution speed, stability and performance.
More information on the findings can be found in this paper.