Next Wednesday, Marcelo Ruas presents his thesis, entitled “Quantile-based time series models for renewable generation probabilistic forecasting”
Marcelo Ruas, researcher at LAMPS and post-graduate student of the Electrical Engineering Department of PUC-Rio, presents his doctoral thesis at 2 PM on Wednesday, April 17, at DEE (room 401 at Cardeal Leme Building).
Producing probabilistic forecasts for renewable generation has become an important topic in recent power systems applications. In this work, the main focus is on generating future scenarios of wind power generation, albeit most of the proposed methods can be directly applied to other renewable sources. In general, most of the time-series methods applied to produce such forecasts rely on the assumption of a conditional Gaussian distribution. However, in practical applications, specially those related with risk analysis, require proper modeling of non-Gaussian processes with focus on distribution tail dynamics. Therefore, in this work, a class of time-series models based on a non-parametric conditional-distribution function is proposed and both estimation and simulation methods are developed. The conditional distribution is composed of an array of connecting jointly-estimated quantile regression models. The first model sets quantiles as a linear function of its covariates with double regularization, namely, one to create a smooth interquantile set of connecting models and another to address covariates selection. The second model is nonparametric function of the covariates, defined a pointwise defined function with controlled second derivative addressed through a filtering scheme. Within the proposed model idea, estimated quantiles are intrinsically connected, creating the idea of a single conditional-distribution model. Case studies with real wind power data from the Brazilian Northeast illustrate the advantages and properties of the models and discussions are conducted.
Alexandre Street (adviser)
Cristiano Fernandes (co-adviser)