Probabilistic short term wind power forecasts using deep neural networks with discrete target classes

摘要

Usually, neural networks trained on historicalfeed-in time series of wind turbines deterministically predictpower output over the next hours to days. Here, the traininggoal is to minimise a scalar cost function, often the root meansquare error (RMSE) between network output and target values. Yet similar to the analog ensemble (AnEn) method, thetraining algorithm can also be adapted to analyse the uncertainty of the power output from the spread of possible targets found in the historical data for a certain meteorologicalsituation. In this study, the uncertainty estimate is achievedby discretising the continuous time series of power targetsinto several bins (classes). For each forecast horizon, a neural network then predicts the probability of power outputfalling into each of the bins, resulting in an empirical probability distribution. Similiar to the AnEn method, the proposedmethod avoids the use of costly numerical weather prediction (NWP) ensemble runs, although a selection of severaldeterministic NWP forecasts as input is helpful. Using stateof-the-art deep learning technology, we applied our methodto a large region and a single wind farm. MAE scores of the50-percentile were on par with or better than comparable deterministic forecasts. The corresponding Continuous RankedProbability Score (CRPS) was even lower. Future work willinvestigate the overdispersiveness sometimes observed, andextend the method to solar power forecasts.