Shallow Neural Networks to Deep Neural Networks for Probabilistic Wind Forecasting

Abstract

The uncertainty associated with wind forecasts is quantified through Neural networks. Comparison between Neural Networks from basic (Feed-Forward) to Deep Neural Networks (Auto-regressive Recurrent Neural Networks) is done. These neural networks are different in architecture as in MLP information flows unidirectionally, in RNN the output of the first time step is fed as input to the next time step whereas in Auto-regressive RNN, parameters are shared between multiple time-series. Auto-regressive RNN learns the trend and seasonality automatically with minimum feature extraction. These methods are used for probabilistic forecasting by addition of projection layer with distribution output. The accuracy and efficiency of these methods are tested on Australian wind power data with 5 min frequency. Prediction intervals with the confidence level of 80%, 85% and 90% are generated through quantiles. These methods prove to be better than other classical probabilistic forecasting methods.