All predictions where performed in R using support vector machine from package e1071 as learner. For each time period, a separate model was tuned. Selection of the optimal model (the SVM hyper-parameters and the selected windows of data) was performed by grammatical evolution (using R package gramEvol) to minimize cross-validation error on the each of the time-series periods.
The first and last time-series were only used for validation of this technique against linear approximation.
The setpoint was discretized to several bins, and a classifier SVM was trained to predict each setpoint based on other 3 time-series.
As a visual inspection of data determined behavior changes around 28th of February, two models were trained for before and after 2014-02-28.
SVM regression was used to predict the first difference (i.e., changes) of the temperature using the other 3 time-series. The first differences were accumulated forwards and backwards, and then averaged to find the true value of temperature.
A similar technique to System supply temperature was used.
A similar technique to Supply temperature setpoint was used, although using only one model.
For each series, a single-step-ahead time-series forecaster was trained based on the previous observed values of all time-series. Both classification and regression were used, as described for individual series.
The forecasting was performed both forwards (as in usual time-series forecasting) and backwards (reverse in time). The resulting predictions were averaged to find the true values.
The daylight saving time was handled separately, using linear interpolation and added later.
Only a "round" function was applied on the "Supply Temperature Setpoint" for the 6th period (all data missing at once). This may introduce a minor rounding error when recomputing the competition score. To compute the exact value as used in the original submission, remove the round function from prediction/pred-5.R line 183.
During the training process, the random number generator is seeded with a constant number. As a result, it is expected that the same models are selected regardless of the execution environment.