Advanced Short-Term Net Load Forecasting for Renewable-Based Microgrids

Abstract

Net load forecasting is essential for the reliable, safe and cost-effective operation of modern power systems incorporating variable renewable technologies. This paper proposes a short-term net load forecasting (STNLF) methodology based on Bayesian neural networks, applicable to microgrids with embedded photovoltaic (PV) systems. Input feature selection and determination of hidden nodes were performed to develop an optimally performing forecasting model. To validate the performance of the model, historical net load site-specific and aggregated data from buildings within the University of Cyprus microgrid (with integrated PV shares of 26%) were used. The developed STNLF model demonstrated a normalized root mean square error of 4.81 % and 3.98% for the solar-integrated building and the microgrid, respectively. Finally, the capability of the developed machine learning forecasting model to yield reliable forecasts was benchmarked against baseline naïve persistence forecasts, achieving skill score improvements of up to 18.61 % at microgrid level.