Effect of automatic hyperparameter tuning for residential load forecasting via deep learning

Abstract

Short-term residential load forecasting is becoming increasingly important as we are advancing to an era where the penetration of renewable energy keeps increasing and will become ubiquitous in our day-to-day energy consumption. In this future grid scenario, individual load forecasting is more critical than load forecasting on system level because many renewable energy sources (RESs) are generally distributed, and it is most efficient to consume the renewable generation at the sites of energy production. Despite there have been many works in short-term load forecasting (STLF), few of them target the problem on the end-use level. Also, deep learning has started to be proposed in STLF, but the common problem for deep learning, the selection of many hyperparameters, is rarely discussed. In this paper, we extend a deep long short-term memory (LSTM) based load forecasting framework with automatic hyperparameter tuning to address the STLF problem for the highly volatile residential load. A tree-structured Parzen estimator based hyper-parameter tuning method is integrated into the STLF framework to efficiently find the best set of hyper-parameters for better forecasting performance.