Prediction techniques adapted to the estimation of energy production of photovoltaic installations integrated in virtual power plants

Resumen

Two irradiance forecasting strategies are distinguished in the thesis. On the one hand, a hybrid day-ahead prediction that is composed of an innovative similar hours-based model together with an artificial neural network-based model, whose predictions are dynamically weighted. At reduced historical datasets, the similar hours-based model produces the highest accuracies, while the neural networks-based model reduces its error over the historical dataset until it outperforms the similar hours-based model, from a four-month historical dataset in the case study. By converting the site into a node within a virtual power plant environment, the similar hours-based model obtains special relevance, since it reduces the error produced to a greater extent compared to other techniques. On the other hand, the second strategy has an intra-day time horizon and it is based on long short-term memory recurrent neural networks. In this case, the shortage of historical data is avoided by using satellite irradiance information, as this is the only information needed by the model. The prediction is updated as new irradiance measurements are obtained at the facility, optimizing its accuracy during the day. Once the irradiance is predicted under different time horizons, an analytical method is used from articles in the literature to determine the photovoltaic power generated in the site, obtaining a reduced error of the method. Afterwards, the power generated at each node in the virtual power plant is evaluated, noticing that the prediction error is considerably reduced and the conclusions observed in the individual nodes are maintained. From the photovoltaic power forecasts, prediction intervals are generated, providing information about the plausible range of photovoltaic power values obtained at the site, for a defined confidence level, considering a Laplacian distribution of the error. The results show that the intervals properly represent the selected confidence level. Once the photovoltaic power produced under different time horizons is estimated, the predictions are unified to generate the optimal power prediction at a defined time instant. To test the potential of the final results, the strategy is employed in a case study of a battery energy management system in the real virtual power plant node in the University. It is observed that the batteries are amortized over their lifetime, allowing the use of these elements without an associated economic loss in a virtual power plant environment, where they are essential.