Performance Comparison of Double Exponential Smoothing and Long Short-Term Memory for Population Forecasting

Abstract

Population forecasting is essential for supporting regional development planning and the effective allocation of public services. However, comparative studies evaluating statistical and deep learning approaches for sub-district population forecasting based on limited historical data are still scarce. This study compares the performance of Double Exponential Smoothing (DES) and Long Short-Term Memory (LSTM) for forecasting the population of 17 sub-districts in Bireuen Regency using annual population data from 2015 to 2024, with forecasts generated until 2027. The dataset was divided chronologically using a time-based split to preserve temporal dependencies, and forecasting accuracy was evaluated using the Mean Absolute Percentage Error (MAPE). Experimental results show that DES achieved lower forecasting errors in most sub-districts, with MAPE values ranging from 1.61% to 4.48%, whereas LSTM produced MAPE values between 1.56% and 7.55%, outperforming DES only in a few sub-districts, such as Peusangan Selatan. In terms of computational efficiency, DES required approximately 0.0007–0.0008 MB of memory and 0.00–1.48 ms processing time, whereas LSTM required 0.0102–0.0118 MB of memory and up to 162.24 ms processing time. The forecasting results indicate a generally increasing population trend across most sub-districts through 2027. These findings suggest that DES is more suitable than LSTM for forecasting annual population data with relatively stable trends and limited historical observations, whereas LSTM may be more appropriate for larger datasets exhibiting more complex temporal patterns. The developed web-based forecasting system provides visualization and forecasting results to support evidence-based population planning in Bireuen Regency.