Estimation and Forecasting of Load: A Comparison of ARIMA, XGBoost and LSTM

With the increasing reliance on electrification and the integration of stochastic renewable energy sources, accurate short-term load forecasting has become critical to ensure the secure and efficient operation of power systems. These developments have introduced greater volatility and non-linearity in electricity demand, posing new challenges for system operators. This study therefore evaluates the performance of three different forecasting models in predicting hourly electricity consumption for Sweden’s Electricity Area 3 (SE3). A statistical model, ARIMA, and two machine learning models, XGBoost and LSTM, were implemented. Historical load data spanning two years, along with temperature as an external covariate, were used for model training and evaluation. All models were developed using the Darts Python library, and their performance was assessed using three standard error metrics: Mean Squared Error (MSE), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). The results demonstrate that the LSTM model achieved the highest forecast accuracy and that none of the three models outperformed the official day-ahead predictions of ENTSO-e's based on the error metrics evaluated. These findings highlight the potential of deep learning and data-driven methods to improve short-term electricity load forecasting in increasingly dynamic power systems.