Monthly electricity consumption prediction using machine learning with LLM-assisted data integration
Abstract
Background . The continuous increase in electricity demand and variability in consumption patterns make it difficult to maintain the supply-demand balance in energy systems, posing significant risks in terms of economic efficiency and sustainability. Therefore, research aimed at accurately prediction energy demand is critical to supporting the efficient management of energy resources and preventing potential supply bottlenecks.
Methods . This study conducted a comprehensive comparative analysis using a total of six algorithms from statistical regression, ensemble learning, and deep learning approaches to model and prediction electricity demand. LLM-based text processing methods were applied to transform data from different sources into a comprehensive dataset. The performance of the models was comprehensively evaluated using the R², Mean Absolute Error (MAE), Mean Squared Error (MSE), and Mean Absolute Percentage Error (MAPE) metrics to assess prediction accuracy and error levels.
Results . The results obtained show that ensemble learning models have higher prediction accuracy and lower error rates. When examining the R² values of model performance, Random Forest and CatBoost stood out as the most effective methods for electricity consumption prediction. However, LSTM and ANN showed superior performance but did not reach the accuracy level of ensemble models. Furthermore, the metric values also support the performance ranking, emphasizing the importance of model selection and hyperparameter optimization on prediction accuracy.