Advanced Time Series Forecasting of Electricity Generation in Turkey: A Comparative Study Using LSTM Models, ARIMA, and PSO - Optimized Holt Trend

Abstract

This study examines energy production forecasting in Turkey through advanced time series methodologies. The analysis, conducted on annual data spanning the period 1970–2024, encompassed not only aggregate electricity generation but also disaggregated production from coal, natural gas, hydro, and renewable sources, including waste. The modeling framework integrated both conventional and machine learning-based techniques, specifically ARIMA, single-layer and three-layer LSTM architectures, and Holt’s linear method optimized via Particle Swarm Optimization (PSO). Model performance was assessed using widely recognized error metrics, namely Mean Absolute Percentage Error (MAPE), Mean Squared Error (MSE), and Root Mean Squared Error (RMSE). The empirical results demonstrated that the three-layer LSTM model achieved the lowest error values for total, coal, and hydro-based generation, whereas the single-layer LSTM exhibited superior accuracy for natural gas. In contrast, the traditional ARIMA model yielded the most precise forecasts for renewable and waste-based energy. These outcomes underscore that while deep learning models such as LSTM are capable of capturing intricate temporal dynamics when appropriately tuned, conventional models like ARIMA continue to demonstrate robust predictive capability for specific datasets. Overall, the findings confirm that all four approaches provide an acceptable level of forecasting accuracy.