On Deep Machine Learning Based Techniques for Electric Power Systems

Sammanfattning: This thesis provides deep machine learning-based solutions to real-time mitigation of power quality disturbances such as flicker, voltage dips, frequency deviations, harmonics, and interharmonics using active power filters (APF). In an APF the processing delays reduce the performance when the disturbance to be mitigated is tima varying. The the delays originate from software (response time delay) and hardware (reaction time delay). To reduce the response time delays of APFs, this thesis propose and investigate several different techniques. First a technique based on multiple synchronous reference frame (MSRF) and order-optimized exponential smoothing (ES) to decrease the settling time delay of lowpass filtering steps. To reduce the computational time, this method is implemented in a parallel processing using a graphics processing unit (GPU) to estimate the time-varying harmonics and interharmonics of currents. Furthermore, the MSRF and three machine learning-based solutions are developed to predict future values of voltage and current in electric power systems which can mitigate the effects of the response and reaction time delays of the APFs. In the first and second solutions, a Butterworth filter is used to lowpass filter the  dq  components, and linear prediction and long short-term memory (LSTM) are used to predict the filtered  dq  components. The third solution is an end-to-end ML-based method developed based on a combination of convolutional neural networks (CNN) and LSTM. The Simulink implementation of the proposed ML-based APF is carried out to compensate for the current waveform harmonics, voltage dips, and flicker in Simulink environment embedded AI computing system Jetson TX2.   In another study, we propose Deep Deterministic Policy Gradient (DDPG), a reinforcement learning (RL) method to replace the controller loops and estimation blocks such as PID, MSRF, and lowpass filters in grid-forming inverters. In a conventional approach it is well recognized that the controller tuning in the differen loops are difficult as the tuning of one loop influence the performance in other parts due to interdependencies. In DDPG the control policy is derived by optimizing a reward function which measure the performance in a data-driven fashion based on extensive experiments of the inverter in a simulation environment. Compared to a PID-based control architecture, the DDPG derived control policy leads to a solution where the response and reaction time delays are decreased by a factor of five in the investigated example.   Classification of voltage dips originating from cable faults is another topic addressed in this thesis work. The Root Mean Square (RMS) of the voltage dips is proposed as preprocessing step to ease the feature learning for the developed  LSTM based classifier. Once a cable faults occur, it need to be located and repaired/replaced in order to restore the grid operation. Due to the high importance of stability in the power generation of renewable energy sources, we aim to locate high impedance cable faults in DC microgrid clusters which is a challenging case among different types of faults. The developed Support Vector Machine (SVM) algorithm process the maximum amplitude and  di/dt  of the current waveform of the fault as features, and the localization task is carried out with  95 %  accuracy.   Two ML-based solutions together with a two-step feature engineering method are proposed to classify Partial Discharges (PD) originating from pulse width modulation (PWM) excitation in high voltage power electronic devices. As a first step, maximum amplitude, time of occurrence, area under PD curve, and time distance of each PD are extracted as features of interest. The extracted features are concatenated to form patterns for the ML algorithms as a second step. The suggested feature classification using the proposed ML algorithms resulted in  95.5 %  and  98.3 %   accuracy on a test data set using ensemble bagged decision trees and LSTM networks.