/IDR-on-GEFCom14

Examine the performance of Isotonic Distributional Regressopn (IDR) (implemented in AlexanderHenzi/isodistrreg) in GEFCom14 forecasting competition

Primary LanguageR

IDR-on-GEFCom14

Goal

Examine performance of Isotonic Distributional Regression (IDR) models (based on CRAN package "isodistrreg", or see https://github.com/AlexanderHenzi/isodistrreg) in GEFCom14 forecasting competition. A comparison with the advanced (and tailored) methods of GEFCom14 participants permits us to get a realistic assessment of the predictive abilities of IDR in probabilistic energy forecasting.

Isotonic Distributional Regression (IDR) [1]

Isontonic Distributional Regression (IDR) was introduced by Henzi et al. as "powerful nonparametric technique for the estimation of conditional distributions under order restrictions" (p.1) [1]. It can be used to estimate conditional cumulative distribution functions (distributional regression!) of a response variable given realizations of explanatory variables and also predict conditional CDFs of the response variable for newly observed explanatory variable values.

IDR on synthetic data

In above plot IDR was applied on bivariate training data. The model provides us with conditional CDF estimates for every x-value in the training data (color gradient is based on the x-value of the training data and helps us to identify respective CDF estimates). Predictions (like the one issued for x=5) correspond to mixtures of estimated conditional CDFs belonging to "close" explanatory variable realizations in the training data.

The key assumption of IDR is isotonicity, which is mathematically defined with partial orders. This order allows us to find "close" explanatory variable values to issue predictions. Essentially, isotonicity means that the response variable assumes larger values with higher probability if the explanatory variables assume larger values. This seems to be given in the training data plot above as the point cloud has an upward trend.

Below, we can compare estimated confidence bands for two synthetic bivariate training data point clouds. The isotonictiy assumption of IDR is crucial to harness its full predictive abilities.

IDR on synthetic data

IDR can easily be used with more than one explanatory variable. For isotonicity, we should look again for variables that impact the response variable positively (negative relations can be inverted) except for variations due to unknown quantities. Theoretically, one could contemplate more in-depth the underlying partial order (for single real-valued explanatory variables the total order on the real values is used), but with the component-wise order there is a useful option predefined. This order has the drawback that the fraction of comparable elements decays exponentially in the number of features.

Global Energy Forecasting Competition 2014 (GEFCom14) [2]

The Global Energy Forecasting Competition 2014 (GEFCom14) was an international and public probabilistic energy forecasting competition. It is covered in Hong et al. [2] and the used data as well as the results of the best contestants, and the leaderboard can be found it the online appendix of this paper.

Participants could take part in solar power, wind power, electricity price and electric load forecasting (the four so-called tracks of GEFCom14). Each track featured 15 succeeding tasks (3 "initial tuning" tasks and 12 evaluated tasks) with incremental data, in which hourly {1,2,3,...,99}% quantiles of the response variable had to be predicted.

Forecasts were evaluated by averaging pinball scores for each task. After dealing with missing and flawed predictions these mean pinball scores per task were translated in skill scores per task for each participant. This was done by calculating the percentual improvements over the mean pinball scores of a benchmark forecast. The final rating corresponds to a linearly weighted average of the 12 skill scores. The last task was weighted 12-times as much as the first task.

Repository

  • simulateCompetition.R:
    Given a prediction and quantile scoring function make predictions with it and calculate scores for all tasks of a given track of GEFCom 14
  • loadData.R, preprocess.R:
    Load and assemble data for a given track and task of GEFCom14 in a unified format, apply transformations after data was loaded (e.g., feature extraction / engineering)
  • applyIDR.R:
    provide wrappers to apply IDR in a variety of different ways
  • util.R:
    Define global variables, help functions, GEFCom14 benchmark predictions and data grouping methods
  • idr_intro.R:
    explore IDR on simulated data and visualize important properties
  • plot.R, exploreData.R, PlotsForThesis:
    visualize and analyze data

Results

IDR can be used competitively for probabilistic energy forecasting. The simple IDR models ("IDR 1") perform already very well. They use just one explanatory variable (current surface solar radiation downwards (solar), wind speed (wind), forecasted load (price), temperature (load)) plus a winter-summer split for the load track. Apart from the wind track the performance can be improved substantially by incorporating more explanatory variables, working with separate models on grouped data (like the winter-summer split), and using subagging (predefined in IDR package).

IDR results in the solar track

IDR results in the wind track

IDR results in the price track

IDR results in the load track

References

  • [1]
    A. Henzi, J. F. Ziegel, and T. Gneiting, "Isotonic distributional regression," Journal of the Royal Statistical Society: Series B (Statistical Methodology), in press, 2021.

  • [2]
    T. Hong, P. Pinson, S. Fan, H. Zareipour, A. Troccoli, and R. J. Hyndman, "Probabilistic energy forecasting: Global Energy Forecasting Competition 2014 and beyond," International Journal of Forecasting, vol. 32, no. 3, pp. 896-913, 2016.