Comparative Analysis of Time Series Forecasting Models for Hourly Energy Demand in Turkey

Abstract

This study conducts an in-depth comparative assessment of five-time series forecasting methods—AutoRegressive (AR), Moving Average (MA), AutoRegressive Moving Average (ARMA), AutoRegressive Integrated Moving Average (ARIMA), and Long Short-Term Memory (LSTM) networks, applied to hourly electricity demand data from Turkey. The objective is to determine which model delivers the highest level of predictive precision and contextual relevance. Model performance was assessed through three key indicators: Mean Absolute Error (MAE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE). The results demonstrate that LSTM exhibits the most favorable accuracy, recording the lowest error values (MAE: 389.21, RMSE: 550.98, MAPE: 1.02%) on the test dataset. This superior performance reflects LSTM’s ability to capture nonlinear behaviors and abrupt demand shifts influenced by external and internal system dynamics. In contrast, conventional models such as AR, MA, and ARMA reported significantly higher forecasting errors, with MAPE values exceeding 14%, indicating limited adaptability to complex and variable consumption patterns. These insights position LSTM as a highly effective tool for improving forecasting reliability, supporting real-time operational planning, and informing strategic energy policies.