Post-processing in solar forecasting: Ten overarching thinking tools

Forecasts are always wrong, otherwise, they are merely deterministic calculations. Besides leveraging advanced forecasting methods, post-processing has become a standard practice for solar forecasters to improve the initial forecasts. In this review, the post-processing task is divided into four categories: (1) deterministic-to-deterministic (D2D) post-processing, (2) probabilistic-to-deterministic (P2D) post-processing, (3) deterministic-to-probabilistic (D2P) post-processing, and (4) probabilistic-to-probabilistic (P2P) post-processing. Additionally, a total of ten overarching thinking tools, namely, (1) regression (D2D), (2) filtering (D2D), (3) resolution change (D2D), (4) summarizing predictive distribution (P2D), (5) combining deterministic forecasts (P2D), (6) analog ensemble (D2P), (7) method of dressing (D2P), (8) probabilistic regression (D2P), (9) calibrating ensemble forecasts (P2P), and (10) combining probabilistic forecasts (P2P), are proposed. These thinking tools can be thought of as the “style” or “mechanism” of post-processing. In that, the utilization of thinking tools circumvents the common pitfalls of classifying the literature by methods (e.g., statistics, machine-learning, or numerical weather prediction), which often leads to a “who used what method” type of roster review that is clearly ineffective, non-exhaustive, and dull. When myriads of post-processing methods are mapped to countable few thinking tools, it allows solar forecasters to enumerate the styles of adjustment that could be performed on a set of initial forecasts, which makes a post-processing task clearly goal-driven. Besides the thinking tools, this paper also emphasizes on the value of post-processing, and provides an outlook for future research. Although this paper is revolved around solar, the materials herein discussed can also be applied to wind and other forecasting areas.