Microgrids of varying size and applications are regarded as a key feature of modernizing the power system. The protection of those systems, however, has become a major challenge and a popular research topic because it involves greater complexity than traditional distribution systems. This paper addresses this issue through a novel approach which utilizes detailed analysis of current and voltage waveforms through windowed fast Fourier and wavelet transforms. The fault detection scheme involves bagged decision trees which use input features extracted from the signal processing stage which are selected by correlation analysis. The technique was tested on a microgrid model developed using PSCAD/EMTDS, which is inspired from an operational microgrid in Goldwind Science Technology Co. Ltd., in Beijing, China. The results showed a great level of effectiveness to accurately identify faults from other non-fault disturbances, precisely locating the fault and trigger opening of the right circuit breaker/s under different operation modes, fault resistances, and other system disturbances.

Accurately forecasting power output of renewable sources is a necessity in operation of today's grid in order to achieve optimal energy utilization and carbon-free ecosystem This study devises a stable, effective and accurate model for dayahead prediction of wind turbine power output through use of an aggregative approach. The method involves two types of Artificial Neural Network (Radial Basis and Conventional Feedforward Networks), Adaptive Neuro-Fuzzy Inference System (ANFIS) and Support Vector Machine (SVM) techniques. It is targeted at comparing the prediction models for their individual performances and finally coming upon an aggregative approach which outperforms the individual models through a strategic combination of them. Three techniques of combining (Simple Averaging, Regression and Outperformance) were tested. Though the individual models showed satisfactory performance by themselves, the combination techniques were able to outperform the individual models. Regression technique of combining was seen to be the most effective of all. The predicted output power through this technique was seen to greatly fit with the measured data with an NMSE of 1.03% for the test year. The combination techniques have also demonstrated more stable performance than the individual models while tested with the extreme cases of windy and less windy weeks.

As the demand and deployment of renewables get better, there comes a growing interest in improved techniques of forecasting the energy generation from those sources. This paper aims at testing and suggesting techniques for input parameter selection and accuracy enhancement in forecasting power output of a PV system. The PV system under study is an operational system in Goldwind smart microgrid in Beijing, China. Historical records of PV power and weather data are utilized while the forecasting models are based on ANN. The paper starts with studying selection of input parameters through correlation analysis, sets of sensitivity analysis techniques and Garson's algorithm. A combination of these methods was able to pick out the most important parameters in deciding the PV power amount. In order to boost accuracy of forecasting, in addition to the recommended strategic input selection method and searching for optimal size of network, options of output processing were tested. The use of more than one network with different training algorithm and using different types of ANN were investigated with both techniques resulting in enhancement in accuracy of forecast.

This article proposes a fault detection and location strategy based on cognitive edge computing to harvest the benefits of cognitive edge computing and address the special needs of active distribution networks (ADNs). In the proposed strategy, an ADN smart gateway is used to compile data in a central repository where it will be processed and analyzed. The intermediary smart gateway includes a protection unit where the fault detection, location, and isolation are accomplished through a combination of virtual mode decomposition (VMD), support vector machine (SVM,) and long short-term memory (LSTM)–type deep machine learning tools. The local measurements of branch currents and bus voltages are processed through VMD, and the informative decomposed components are provided as inputs to the SVM-based fault detection unit and LSTM-based fault location unit. The smart digital relay passes trip commands to the respective circuit breaker/s and submits compiled data regarding the history of faults and protection actions to the upper-level units. The findings from simulation results demonstrate the effectiveness of the proposed strategy to provide fast and accurate fault detection and protection against all types of faults and locations in the ADN.

Short term Load Forecasting (STLF) has become an important feature of the `smart grid' concept. Though there has been considerable efforts, especially in the aggregated load aspect, luck of a clear-cut recommendation on single method of forecasting due to the high irregularity of load patterns and variation of load characteristics of different consumer types has made it still a topic that needs further study. In this study, an approach composed of wavelet transform and autoregressive neural networks is devised and tested to prove an improved forecast for single commercial type customer. The method proposes breaking down the load data into an approximate and detail components through wavelet decomposition that will be treated and forecasted independently. A unique step of determining the optimal feedback delays for each target parameters through autocorrelation analysis is adopted. A cross correlation analysis is employed to decide on the usage of external input parameters and hence choose either of the NARnet or NARXnet models. The individual forecasts are later merged trough wavelet reconstruction to give the load forecast. The technique was tested using practical load and weather data. The method showed high effectiveness compared to NAR and NARX networks as well as SVM and ANFIS techniques.