DATA SCIENCE-BASED MODELING FOR ELECTRICITY CONSUMPTION FORECASTING

Abstract

Accurate forecasting of electricity consumption is central in modern energy systems for both planning and operational decision making. This research proposes a data science driven methodology for predicting electricity demand and directly compares machine learning with deep learning approaches under identical experimental conditions. Specifically, a tree based model, XGBoost, and a recurrent neural network, Long Short-Term Memory (LSTM), have been evaluated on the same dataset with consistent data preprocessing, splits, and metrics. Currently, tree based models use explicit temporal feature engineering in order to find nonlinear relationships in structured time dependent data. In contrast, LSTM networks learned the temporal dependencies from sequential load observations directly. The performance evaluation of these models is made by RMSE and MAPE. Accordingly, it follows that LSTM yields the highest forecasting accuracy with the superior capturing of temporal dynamics, while XGBoost offers competitive performance with lower computational complexity. This implies that there are some important context-dependent tradeoffs that exist among model accuracy, interpretability, and efficiency when coming to electricity consumption forecasting.