Modelling and Forecasting of Photovoltaic Generation for Microgrid Applications: from Theory to Validation

The penetration of stochastic renewable generation in modern power systems requires to reconsider conventional practices to ensure the reliable functioning of the electrical network. Decentralized control schemes for distributed energy resources (DERs) have gained attention to support the grid operation. In order to cope with the uncertainties of the DERs, predictive schemes that leverage on forecast of renewable generation recently came into prominence. The period of the control action typically depends on the availability of the reserve in the grid. For the case of microgrids, their limited physical extension and the lack of spatial smoothing imply fast power fluctuations and the necessity of coupling energy management strategies with real-time control. Among the DERs, small-scale photovoltaic (PV) systems are expected to represent most of the future available capacity, and consequently, solar resource assessment and power forecasting are of fundamental importance. This thesis focuses on developing forecasting methods and generation models to support the integration of photovoltaic systems in microgrids, considering short-term temporal horizons (below one hour) and fine spatial resolution (single site installations). In particular, we aim at computing probabilistic prediction intervals (PIs), i.e. we include information accounting for the intrinsic uncertainty of the prediction. In this respect, nonparametric tools to deliver PIs from sub-second to intra-hour forecasting horizons are proposed and benchmarked. They forecast the AC power and/or the global horizontal irradiance (GHI) by extracting selected endogenous influential variables from historical time series. The methods are shown to outperform available state-of-the-art techniques, and are able to capture the fastest fluctuations of small-scale PV plants. Then, we investigate how the inclusion of features from ground all-sky images can be used to improve time-series-based forecasting tools, thanks to identifying clouds movement. In this respect, we define a toolchain that allows predicting the cloud cover of the sun disk, through image processing and cloud motion identification. Furthermore, a methodology to estimate the irradiance from all-sky images is proposed, investigating the possibility of using an all-sky camera as an irradiance sensor. Next, we consider the problem of having power measurements that are corrupted by exogenous control actions (e.g. curtailment) and, therefore, not representative of the true potential of the PV plant. We propose a model-based strategy to reconstruct the maximum power production of a PV power plant thanks to integrating measurements of the PV cell temperature, system DC voltage and current. The strategy can improve time series-based direct power forecasting techniques when the production of the PV system is curtailed and thus the measured power does not correspond to the maximum available. The proposed methods to model and forecast the PV generation are then integrated in a single chain that allows to deliver power PIs that are able to account for the overall uncertainty of a PV system at a predefined confidence level. In the last part of the thesis, the proposed methods are experimentally validated in a real microgrid by considering possible applications in modern power systems.