Using machine learning to enhance skill of subseasonal-to-seasonal (S2S) temperature forecasts

Energy demand, especially for residential heating, is largely driven by temperature.  High-quality subseasonal-to-seasonsl (S2S) forecasts of temperature are therefore valuable for risk management and energy trading, yet the use of these forecasts is often limited by their complexity, by difficulties in combining different forecast types, and by the relatively weak probabilistic skill they produce. Sequential learning algorithms (SLA), offer a means to overcome many of these difficulties.  SLAs optimally combine information from multiple ‘experts’ or predictors using weights and reduce forecast bias by continuously learning over time from each forecast verification.  The information from these ‘experts’, which can both be statistical or numerical forecasts, are used as SLA inputs and blend into a single information stream through dynamical updating of the weights. Here, ‘experts’ are defined as quantiles of raw ECMWF S2S 2 m temperatures (T2m) forecasts without bias adjustment and ERA5 T2m climatology. The SLA produces probabilistic forecasts of Great Britain-averaged T2m at lead times of 1-4 weeks for the period 2004-2023.  Results show positive anomaly correlation co-efficient and rank probability skill scores for SLA T2m forecasts across all weeks compared to both the raw S2S and climatological forecast.  Analysis of the weight evolution shows that SLA relies heavily on the raw forecast experts at weeks 1-2 but shifts towards climatological experts in the later weeks, with a clear seasonal evolution to the weight profile. It is also confirmed that this online-learning approach with adaptive weights outperforms the most optimal static weight combination even though the latter is permitted the benefit of perfect foresight.