A new approach to extended‐range multimodel forecasting: Sequential learning algorithms

Abstract

Multimodel combinations are a well-established methodology in weather and climate prediction and their benefits have been widely discussed in the literature. Typical approaches involve combining the output of different numerical weather prediction (NWP) models using constant weighting factors, either uniformly distributed or determined through a prior skill assessment. This strategy, however, can lead to suboptimal levels of skill, as the performance of NWP models can vary with time (e.g., seasonally varying skill, changes in the forecasting system). Moreover, standard combination methods are not designed to incorporate predictions derived from sources other than NWP systems (e.g., climatological or time-series forecasts). New algorithms developed within the machine learning community provide the opportunity for "online prediction" (also referred to as "sequential learning"). These methods consider a set of weighted predictors or "experts" to produce subsequent predictions in which the combination or "mixture" is updated at each step to optimize a loss or skill function. The predictors are highly flexible and can combine both NWP and statistically derived forecasts transparently. A set of these online prediction methods is tested and compared with standard multimodel combination techniques to assess their usefulness. The methods are general and can be applied to any model-derived predictand. A set of weather-sensitive European country-aggregate energy variables (electricity demand and wind power) is selected for demonstration purposes. Results show that these innovative methods exhibit significant skill improvements (i.e., between 5 and 15% improvement in the probabilistic skill) with respect to standard multimodel combination techniques for lead weeks up to 5. The incorporation of statistically derived predictors (based on historical climate data) alongside NWP forecasts is also shown to contribute significant skill improvements in many cases.

The Role of Weather Predictions in Electricity Price Forecasting Beyond the Day-Ahead Horizon

“…Hence, past performance leads, at times, to a suboptimal model combination for forecasting purposes. This finding complements previous studies that conducted experiments in more stable market conditions [4], [42], [44], [46]. As discussed in the previous Section, a further drawback of the Combined forecasts is that weather forecasts are a shared source of errors, thus resulting in error correlation between forecasts.…”

A Multi‐Model Ensemble Kalman Filter for Data Assimilation and Forecasting

“…Sengupta et al (2020) used a Bayesian neural network to infer model weights. Sequential aggregation takes inspiration from online learning and game theory in weighting forecasts with rules that have theoretical performance guarantees (Gonzalez et al, 2021;Mallet et al, 2009;Thorey et al, 2017). Forecast weights can also be determined using Markov chain Monte Carlo (Dumont Le Brazidec et al, 2021).…”

Smoothed Bernstein Online Aggregation for Short-Term Load Forecasting in IEEE DataPort Competition on Day-Ahead Electricity Demand Forecasting: Post-COVID Paradigm

“…The BOA will lead always to a convex combination of the forecasters, as the EWA. Further, is well known that the EWA in combination with the gradient trick can achieve optimal convergence rates, if the considered updating loss is exp-concave, see [29], [30]. Unfortunately, the required absolute deviation AD is not expconcave.…”