Short-Term Load Forecasting using Long Short Term Memory Optimized by Genetic Algorithm

Abstract

In the routine operation of a smart grid (SG), accurate short-term load forecasting (STLF) is paramount. To predict short-term load more effectively, this paper proposes an integrated evolutionary deep learning strategy based on navel feature engineering (FE), long short-term memory (LSTM) network, and Genetic algorithm (GA). First, FE eradicates repetitious and irrelevant attributes to guarantee high computational efficiency. The GA is then used to optimize the parameters (ReLU, MAPE, RMSprop batch size, Number of neurons, and Epoch) of LSTM. The optimized LSTM is used to get the actual STLF results. Furthermore, most literature studies focus on accuracy improvement. At the same time, the importance and productivity of the devised model are confined equally by its convergence rate. Historical load data from the independent system operator (ISO) New England (ISO-NE) energy sector is analyzed to validate the developed hybrid model. The MAPE of the proposed model has a small error value of 0.6710 and the shortest processing time of 159 seconds. The devised model outperforms benchmark models such as the LSTM, LSTM-PSO, LSTM-NSGA-II, and LSTM-GA in aspects of convergence rate and accuracy. In other words, the LSTM errors are effectively decreased by the GA hyperparameter optimization. These results may be helpful as a procedure to shorten the time-consuming process of hyperparameter setting.