Short-Term Load Forecasting Employing Recurrent Neural Networks

This research paper explores the domain of Short-Term Load Forecasting (STLF) through the application of Long Short- Term Memory (LSTM) and Gated Recurrent Unit (GRU) models. Focused on the west region of the Electric Reliability Council of Texas (ERCOT), the study employs a dataset spanning 2012–2015, enriched with new features such as day of the month, month of the year, and a detailed holiday classification. The dataset also incorporates hourly temperature, humidity, time of the day, day of the week, and hourly load values, offering a comprehensive perspective for load forecasting. Through meticulous hyperparameter tuning, LSTM and GRU models are optimized to find their best architecture. Four evaluation metrics have been employed to assess the performance of our models, namely, Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the coefficient of determination or R Squared (R2). With the identified optimal hyperparameters and look-back window, the obtained values for MAE, RMSE, MAPE and R2 for the LSTM model are 8.94, 12.24, 0.80 and 0.9973, respectively. The GRU model also achieves good performance with MAE, RMSE, MAPE, and R2 of 10.72, 14.48, 0.96, and 0.9962, respectively. Our research shows that the accuracy of STLF on the dataset has been improved by adding new features and fine-tuning the LSTM and GRU models, with the LSTM model achieving the best performance.