Data driven modelling of distribution system considering a high penetration of renewable energy sources for control applications

The distribution system is undergoing a transition process of modernisation, where it is expected to make more efficient use of the available resources and equipment. This creates challenges for maintaining the system operating within the allowed operational conditions considering the constant changes of “bidirectional” power flows. To tackle these challenges, the use of available data may help to describe the distribution system without relaying on “snapshots” that represent worst-case scenarios or non-updated information from any of the traditional electric parameters used in modelling. This thesis focused on the definition of an approach to analyse and model the voltage in distribution systems with high penetration of renewable energy, and the (potential) use of these models on control applications. A review of the relevant data-driven modelling approaches was conducted, including the background of power systems parameters and modelling, voltage modelling and control for distribution systems with high penetration of renewables, and observability and controllability in distribution systems. Different modelling approaches in distribution systems were reviewed, to consider the potential inclusion of measurable data and new metrics to help detect the desired system dynamics to be represented. This data is a set of time-series measurements that are expected to describe the distribution system itself. An innovative description of the distribution system was introduced, by using a set of new proposed metrics. These metrics were based on power (dissipated) in lines provided and voltage covariance between nodes to describe, respectively, size and distance of perturbations. They give relevant spatial-temporal information of the distribution system and its perturbations based on time-series data. Different scenarios were explored to evaluate the limits of the amount of information that can be extracted from the distribution system. Results showed that the metrics can represent spatial-temporal features and events occurring in the distribution system, which make them suitable for real-time applications. Once the metrics were obtained, an algorithm was proposed to produce a linear model that represents the distribution system using measurable data. This algorithm requires a revision of data to understand the structure of the proposed model. In this case, the time-series data must be stationary to produce the linear model. Once this condition was achieved, the next step was to produce a provisional model to explore the proposed regressors used in modelling. Once a reference case was obtained, the input data was analysed to detect any conditions that may introduce errors in the model (e.g., collinearity in some exogenous regressors), improve the response (e.g., analysis of cross-validation using Granger-causality) and highlight the regressors and time lags that improve the response. The first and final models were compared to explore if the proposed metrics could explain the system dynamics. Results after applying the algorithm showed that it was possible to obtain a good model for one-step ahead prediction, which can be easily integrated to any control structure. Finally, the refined model was improved by presenting a prediction interval based on bootstrapping and cross-validation techniques used in time-series data.