Classes for creating forecasting models of univariate time series and time series with exogenous variables
Motivation • Description • Supported algorithms • Credits • How To Use • Data used
Each individual implementation of the prediction algorithm from Darts has its own grid_search method, but they do not allow searching multiple models at the same time. The developed class is a wrapper over the most interesting prediction methods from this library.
- The GridSearch class allows to obtain the optimal predictive model for certain data based on specified algorithms and a grid to search parameters for each.
- The Imputation class is the interface to the basic classical methods for imputing time series misses:
- Mean, median and mode
- Previous, next and average of the nearest to the missing value
- Polynomial function
- Spline
- Moving average
GridSearch is currently supported for the following algorithms (as the most interesting ones):
- RNNModel
- CatBoostModel
- TFTModel
- TCNModel
An example of using GridSearch classes is given in the module file itself, an example of using Imputation is given in the imputation_usage.py file
The project was created to predict the series of electric and heat energy consumption.
Here you can access the data used.
Notebook with preparation and analysis of the initial dataset