Advanced forecasting methods for renewable generation and loads in modern power systems

The PhD Thesis deals with the problem of forecasting in power systems, i.e., a wide topic that today covers many and many needs, and that is universally acknowledged to require further deep research efforts. After a brief discussion on the classification of forecasting systems and on the methods that are currently available in literature for forecasting electrical variables, stressing pros and cons of each approach, the PhD Thesis provides four contributes to the state of the art on forecasting in power systems where literature is somehow weak. The first provided contribute is a Bayesian-based probabilistic method to forecast photovoltaic (PV) power in short-term scenarios. Parameters of the predictive distributions are estimated by means of an exogenous linear regression model and through the Bayesian inference of past observations. The second provided contribute is a probabilistic competitive ensemble method once again to forecast PV power in short-term scenarios. The idea is to improve the quality of forecasts obtained through some individual probabilistic predictors, by combining them in a probabilistic competitive approach based on a linear pooling of predictive cumulative density functions. A multi-objective optimization method is proposed in order to guarantee elevate sharpness and reliability characteristics of the predictive distribution. The third contribute is aimed to the development of a deterministic industrial load forecasting method suitable in short-term scenarios, at both aggregated and single-load levels, and for both active and reactive powers. The deterministic industrial load forecasting method is based on multiple linear regression and support vector regression models, selected by means of 10-fold cross-validation or lasso analysis. The fourth contribute provides advanced PDFs for the statistical characterization of Extreme Wind Speeds (EWS). In particular, the PDFs proposed in the PhD Thesis are an Inverse Burr distribution and a mixture Inverse Burr – Inverse Weibull distribution. The mixture of an Inverse Burr and an Inverse Weibull distribution allows to increase the versatility of the tool, although increasing the number of parameters to be estimated. This complicates the parameter estimation process, since traditional techniques such as the maximum likelihood estimation suffer from convergence problems. Therefore, an expectation-maximization procedure is specifically developed for the parameter estimation. All of the contributes presented in the PhD Thesis are tested on actual data, and compared to the state-of-the-art benchmarks to assess the suitability of each proposal