Forecasting Load Profiles and Critical Overloads with Uncertainty Quantification for Low Voltage Smart Grids

Abstract

Forecasting load profiles and critical overloads is essential to ensure the reliability and safety of Low Voltage Smart Grids. We specifically target to forecast the total grid load, and partial loads by applying an LSTM recurrent neural network combined with quantile uncertainty quantification. The design of the proposed forecasting and uncertainty prediction algorithms allows for tuning the sensitivity of the overload detection and for the deployment to low-cost edge devices. The forecasting algorithms were tested with one-year simulations across six different Smart Grid topologies. Our results demonstrate consistent and robust forecasting performance across various simulations. The overload detection capabilities are analyzed based on a single simulation scenario and show the effectiveness of using different quantile prediction intervals to tune the sensitivity of the overload detection performance.