Quantile Combination for the EEM20 Wind Power Forecasting Competition

Abstract

Combining forecasts is an established strategy for improving predictions and is employed here to produce probabilistic forecasts of regional wind power production in Sweden, finishing in second place in the EEM20 Wind Power Forecasting Competition. We combine quantile forecasts from two models with different characteristics: a 'discrete' tree-based model and 'smooth' generalised additive model. Quantiles are combined via linear weighting and the resulting combination is superior than both constituent forecasts in all four regions considered.

Probabilistic load forecasting for the low voltage network: Forecast fusion and daily peaks

“…If the conventional forecast were a prediction of load conditional on there not being a peak at d, h, this could be interpreted as a statement of the law of total probability. While it may be possible to produce such a forecast, we choose to proceed using conventional day-ahead forecasts so that the fusion method is as applicable as possible, Forecast fusion is related to techniques found in probabilistic forecast combination [22,23], blending [24], expert mixtures [25], linear pools [26,27], and so on. Empirically it was found that important concepts in forecast combination, such as re-calibration of a linearly combined forecasts, were not necessary in the following case study.…”

Forecast combinations: an over 50-year review

“…One line of research has looked at tailoring the individual weights for different quantile levels, i.e., a separate weight is allocated for each individual model and each quantile level, by replacing w T+h|T,i with w T+h|T,i (τ) in Equation (20). For example, individual quantiles can be weighted by the reciprocal of the value of the pinball loss function (Wang et al, 2019;Zhang et al, 2020;Browell et al, 2020). This flexible strategy enables the combination to accommodate the fact that individual forecasting models may have varying performance at different quantile levels.…”

“…One exception is who looked at the statistical properties of the simple average of probability forecasts and the ability to benefit from averaging. The choice of the combination weights has only been explored empirically, mainly in the context of energy forecasting (e.g., Wang et al, 2019;Browell et al, 2020) and epidemiological forecasting (e.g., Brooks et al, 2020;Ray et al, 2020). Some of these proposals appear practical and beneficial, while some of them appear less useful.…”

Probabilistic load forecasting for the low voltage network: forecast fusion and daily peaks

“…If the conventional forecast were a prediction of load conditional on there not being a peak at d, h, this could be interpreted as a statement of the law of total probability. While it may be possible to produce such a forecast, we choose to proceed using conventional day-ahead forecasts so that the fusion method is as applicable as possible, Forecast fusion is related to techniques found in probabilistic forecast combination [19,20], blending [21], expert mixtures [22], linear pools [23,24], and so on. Empirically it was found that important concepts in forecast combination, such as re-calibration of a linearly combined forecasts, were not necessary in the following case study.…”

“…Forecast evaluation is reported via the relative change of the score of the proposed model S to a benchmark S re f via skill scores. If the perfect score is zero, as in the cases considered here, then the skill score is skill = 1 − S S re f (19) and in the following the terms skill score, percentage improvement, and relative change are used interchangeably. Bootstrap re-sampling is used as a simple non-parametric method for estimating the significance in forecast improvement [37].…”