Model 1.5.1
- Citation Author(s):
-
Konstantinos Theodorakos
- Submitted by:
- Konstantinos Theodorakos
- Last updated:
Abstract
Model 1.5.1
Dropout Additive Regression Trees (DART) + Clustered-meter average monthly ratios
Monthly (kWh) consumption clustering:
1. Twelve clustering models (one per month of signup): Spectral Clustering
2. Clustering using (daily kWh) consumption extracted features. Cluster count set to 4 (empirically better than the 5 or 6 suggested by the elbow distortion method).
Forecasting for a meter: Using the mean monthly consumption + in-cluster month-to-month ratios.
Clustering
Features (170) from weekend/weekday and full series:
- Statistical: median, variance, quantiles, ...
- Time-series: autocorrelations, trends, seasonalities, ...
Month-to-month ratios:
- Robust STL LOESS trend was used to calculate the month-to-month percentage ratios. All monthly averages are replaced with the new computed ratios.
Preprocessing:
- Time-series: Converted 0 to Nans and dropped them.
Regression Forecasting:
- Using DART Gradient Boosted Trees (GBT) on the calculated 2018 month ratios + Nested Cross-validation. Dual Annealing used for hyper-parameter optimization.
