Development of a distributed machine learning platform with feature augmented attributes for power system service restoration
Modern power systems are gradually adopting the philosophy of autonomous and distributed means of dynamic event detection processes, by facilitating system operation using intelligent algorithms. Dynamic data has different forms, such as voltage, current, active-reactive power, frequency, rotor speed and angle, status of a breaker, electricity demand, electricity price, operation schedules and many more. Collecting and storing these dynamic data from a system has recently become much more feasible for engineers. Along side that, computational resources have also been increased in multitude. Implementing these resources can bring practical benefits for post-event operation under uncertain conditions. The market drive towards decentralization and deregulation is constantly introducing problems which are stochastic. Thus, techniques like machine learning are observing a fair share of its applications in the field of power engineering. Though in recent years, application of machine learning algorithms in the power industry has become a prominent domain, understanding a dynamic data set requires additional features. Developing such features is yet an interesting research genre where significant contributions can be made. The concept of feature analysis using power system dynamic data is relatively novel.
This research, in multiple steps, applies different machine learning methods to understand the dynamic nature of power system data. The idea is to search for effective methods to analyse this power system data. The dynamic nature of any power system demands efficient training and testing facilities in a pattern recognition system. Thus, recognizing the impacts of different dynamic techniques on a continuous flow of power system data would be a priority in this work and it requires a prelude to realizing the capacity of data analytics to understand power system dynamics.
However, investigating the dynamic capabilities of the machine learning techniques is only the beginning. The power system is a complex network that has lots of segments and interconnections, with predefined priorities assigned. Multiple pattern recognition techniques may, therefore, be required to effectively understand the interactions between events that govern system operation and control. Rather than analysing the impact of a method on all types of scenarios, preparing different effective methods for different events would be a logical approach. The question arises whether a centralized scheme can be of any use while understanding events that are occurring in different autonomous zones. For example, the dynamic behavior of a system near a wind farm, reacting to different wind power penetrations, could effectively be understood with a decentralized system that is already installed in a nearby substation. Under such a scenario, collecting all the system data and storing it to the central station for further analysis should be computationally expensive, considering that some events are taking place at that very moment. It is another motivating factor for this work to investigate whether distributed data analysis with dynamic data handling capacity can lead towards better prediction, and consequently, the operation of a system.
However, power system dynamics and machine learning algorithm-based data analysis are two independently broad domains. Analysing the impacts of data analytics on power system dynamics, would therefore be another broad discussion, ranging from a generic towards a specific study. In this thesis, machine learning platforms have been considered as the intelligent means of data analysis and the self-healing framework of a microgrid has been used for classification of the dynamic power system events. The key contribution of this research is developing a feature-extraction-based algorithm that has the capacity of detecting power system events and facilitating decision making.