A Comparison of Univariate Methods for Day-Ahead Short-Term Load Forecasting

Abstract

Short-Term Load Forecasting (STLF) plays a crucial role in energy management for planning and managing the operation strategy of hybrid energy sources. A common STLF approach involves making predictions based on multiple input variables which affect the load. However, additional data besides historical load data often are not available, making STLF a challenging univariate task. Two approaches are discussed in this paper for the univariate day-ahead STLF: the first integrates temporal information into the input regression vector, with Seasonal Auto-Regressive Moving Average (SARMA) and Prophet being the best representatives for the application considered; the second integrates temporal information directly into its architecture through the feedback of the internal state, with Long Short-Term Memory (LSTM) and vanilla Recurrent Neural Network (RNN) chosen for this purpose. The models were evaluated on a real case study involving a university building in Italy. The results validate the effectiveness of the LSTM model. LSTM performed significantly better than SARIMA, and slightly better than RNN and Prophet, achieving a Mean Absolute Error (MAE) equal to 13.12 kW, a Root Mean Square Error (RMSE) equal to 25.09 kW, a Mean Absolute Percentage Error (MAPE) equal to 10.97% and an R2 score equal to 0.85. This work provides a foundation for the improvement and deployment of such time series forecasting techniques in the emerging field of STLF for energy management.