Mixture Density Networks applied to wind and photovoltaic power generation forecast
Damian Vallejo, Ruben Chaer
Abstract
In this work, the training of a Mixture Density Network (MDN) type of Neural Network (NN) is presented. This network is used to forecast the power generated by wind and photovoltaic farms in Uruguay in a one week time frame. With the MDN model, not only the expected value of the hourly power generation is forecasted, but also a probability density function for each signal. This allows to provide information not only about the expected value of the power forecasted but also for how certain this value is estimated to be. The inputs of the network are meteorological values acquired from a private vendor and the output is the power generation probability density function. A comparison between the previously used models and the new one is shown and future improvements are discussed.