An ultra-fast way of searching weather analogs for renewable energy forecasting

摘要

Analogs-weather patterns that highly resemble each other-have been widely adopted by the meteorology and renewable energy communities for predictive applications. It has been repeatedly demonstrated that by searching for and using past analogs, an analog ensemble (AnEn) can be constructed, circumventing or complementing computationally expensive dynamical ensemble systems. Of course, the pattern matching required by the AnEn benefits from larger historical datasets. However, brute-force analog searches become impractical with larger training datasets. To overcome this challenge, this paper introduces a rapid method for finding analogs. This method is referred to as Mueen's algorithm for similarity search (MASS). MASS has many desirable properties in that it is exact, non-parametric, scalable, parallelizable and most notably, free from the curse of dimensionality. MASS can also be easily extended to multivariate cases. In a case study, 20 years of 1-h averaged ground-based multivariate weather data are used to exemplify a typical AnEn setup. It is found that MASS is about 100 times faster than the brute-force algorithm. MASS is suitable for all Euclidean distance-based pattern-matching tasks.