Spatio-temporal forecasting and optimization for integration of solar energy in urban energy systems

The increasing penetration of non-dispatchable renewable energy sources such as photovoltaic (PV) systems in the electricity generating mix poses challenges to the operational performance of the power system. On the demand side, advanced schemes that increase the flexibility of customer loads and the increase in electrification are set to noticeably alter electricity demand. Moreover, a rooftop-mounted PV system alters the electricity demand of the building it is connected to because the generated electricity first serves the electrical loads of the building, thereby affecting the so-called net load that the electricity grid experiences. This thesis studies solutions that enhance the integration of dispersed solar PV power into the power system, with a particular focus on probabilistic and multivariate forecasts and a control framework based on such forecasts. In addition, the thesis evaluates voltage control by means of reactive power control of solar PV inverters. Probabilistic solar, load and net load forecasts are generated using static and dynamic forecast models, where the latter result in approximately 99% less computation time and improved calibration and sharpness but lower forecast resolution. The dynamic forecast models are subsequently used to study the impact of spatial aggregation of customers on the predictive densities, which results in improved calibration and sharpness. Interestingly, the positive effects are already noticeable when aggregating a few customers and this can lead to improved decision-making on the community level. Multivariate forecasts in the form of time and space-time trajectories are also studied, where the multivariate distribution is represented by a copula. Specifically, it is shown that an empirical copula is particularly suitable for high-dimensional spatio-temporal forecasts whereas a Gaussian copula is well suited for temporal forecasts with a large forecast horizon. Furthermore, the thesis develops an augmented version of the scenario-based stochastic model predictive control algorithm that implements the global optimal control action---if it exists---rather than the expectation of independent optimal control actions, which manages forecast errors more effectively. Finally, a population based search method is applied to reactive power control that is able to explicitly and independently model the spatial and temporal relationship between dispersed solar PV inverters, resulting in an improved voltage profile with a smaller population compared to benchmarks. In summary, the thesis demonstrates that forecasts can be improved using several methods, e.g., by spatially aggregating customers, by combining PV power generation and electricity use, by preselecting informative predictors or by postprocessing the forecasts. In turn, the improved accuracy of the forecasts can increase their value in applications such as optimal control problems, that can improve solar PV integration in urban energy systems