Protection-based Distributed Generation Penetration Limits on MV feeders - Using Machine Learning

The rise of disruptive technologies and the rapid growth of innovative initiatives have led to a trend of decentralization, deregulation, and distribution of regulated/centralized services. As a result, there is an increasing number of requests for the connection of distributed generators to distribution networks and the need for power utilities to quickly assess the impacts of distributed generators (DGs) to keep up with these requests. Grid integration of DGs brings about protection issues. Current protection systems were not designed for bi-directional power flow, thus the protective devices in the network lose their ability to perform their main functions. To mitigate the impact of distributed generation (DG), some standards and policies constrain the number of DG that can be connected to the distribution network. The problem with these limits is that they are based only on overload and overvoltage, and do not adequately define the DG size/threshold before the occurrence of a protection issue (NRS 097-2-3). The other problem with distributed generation is the vast difference in the technology, location, size, connection sequence, and protection scheme requirements which results in future DG network planning inadequacies – The Network DG Planning Dilemma. To determine the amount of DG to connect to the network, a detailed analysis is required which often involves the use of a simulation tool such as DIgSILENT to model the entire network and perform load flow studies. Modelling networks on DIgSILENT is relatively easy for simple networks but becomes time-consuming for complex, large, and real networks. This brings about a limitation to this method, planning inadequacies, and longer connection approval periods. Thus, there is a need for a fast but accurate system-wide tool that can assess the amount of DG that can be connected to a network. This research aims to present a technique used for calculating protection-based DG penetration limits on MV networks and develop a model to determine medium voltage opportunity network maps. These maps indicate the maximum amount of DG that can be connected to a network without the need for major protection scheme changes in South Africa. The approach to determining protection-based penetration limits is based on supervised machine learning methods. The aim is to rely on protection features present in the distribution network data i.e. fault level, Inverse Definite Minimum Time (IDMT) curve, pick-up current settings, Time Multiplier Settings (TMS), calculated relay operating times and relay positions to see how the network responds at certain DG penetration levels (‘actual' relay operating times). The dataset represents carefully anonymized distribution networks with accepted protection philosophy applied. A supervised machine learning algorithm is applied after nontrivial data pre-processing through recommendation systems and shuffling. The planning dilemma is cast into three parts: the first part is an automated pattern classification (logistic regression for classification of protection miscoordination), the second part involves regression (predicting operating time after different levels of DG penetration), and the last part involves developing a recommendation system (where, when and how much photovoltaic (PV) DG will be connected). Gradient descent, which is an optimisation algorithm that iterates and finds optimal values of the parameters that correspond to the local or global minimum values of the cost function using calculus was used to measure the accuracy of each model's hypothesis function. The cost function (one half mean squared error) for the models that predict ‘actual' relay operating times before DG penetration, at 35%, 65%, and 75% DG penetration converged to values below 120, 20, 15, and 15 seconds2 , respectively, within the first 100 iterations. A high variance problem was observed (cross-validation error was high and training error was low) for the models that used all the network protection features as inputs. The cross-validation and training errors approached the desired performance of 0.3±0.1 for the models that had second-order polynomials added. A training accuracy of 91.30%, 73.91%, 82.61%, and a validation accuracy of 100%, 55.56%, 66.67% was achieved when classifying loss of coordination, loss of grading and desensitization, respectively. A high bias problem was observed (cross-validation error was high and training error was high) for the loss of grading classification (relay positions eliminated) model. When the models (horizontal network features) were applied to four MV distribution networks, loss of coordination was not predicted, the loss of grading model had one false positive and the de-sensitization model had one false negative. However, when the results were compared to the vertical analysis (comparing the operating times of upstream and downstream relays/reclosers), 28 points indicated a loss of coordination (2 at 35%, 1 at 65% and 25 at 75% DG penetration). Protection coordination reinforcements (against loss of grading and desensitization) were found to be a requirement for DG connections where the MV transformer circuit breaker TMS is between 0.5 and 1.1, and where the network fault level is between 650 and 800A. Distribution networks in affluent neighbourhoods similar to those around the Western CapeSomerset West area and Gauteng- Centurion area need to be reinforced to accommodate maximum DG penetration up to the limit of 75% of the After Diversity Maximum Demand (ADMD). For future work, the collection of more data points (results from detailed analytical studies on the impact of DG on MV feeders) to use as training data to solve the observed high variance problem is recommended. Also, modifying the model by adding upstream and downstream network features as inputs in the classification model to solve the high bias problem is recommended