Rock aggregate demand forecasting using long short-term memory optimized by particle swarm optimization

Source abstract

Background: Accurate rock aggregate demand forecasting is essential for logistics planning and fleet allocation, yet highly fluctuating demand, missing observations, and non-linear temporal patterns can challenge conventional forecasting approaches. Aim: This study aimed to evaluate a computational forecasting approach using Long Short-Term Memory optimized by Particle Swarm Optimization (LSTM-PSO) and compare its predictive performance with Croston Syntetos-Boylan Approximation (SBA) and Auto-ARIMA. Method: Monthly rock aggregate demand from PT. Hasal Logam Utama covering January 2017–December 2021 was analyzed. Seven missing observations were treated using seasonal imputation and linear interpolation, and the resulting 60-month series was divided into 54 training and six testing observations. PSO optimized five LSTM hyperparameters, while model performance was assessed using sMAPE, MASE, RMSSE, and Forecast Accuracy (FA). Results: LSTM-PSO achieved an sMAPE of 11.83%, MASE of 0.1552, RMSSE of 0.1139, and FA of 88.12%, compared with Croston SBA (sMAPE=17.71%; FA=82.93%) and Auto-ARIMA (sMAPE=16.86%; FA=82.15%). Conclusion: LSTM-PSO demonstrated promising forecasting performance for the examined rock aggregate demand series, although validation using longer datasets and additional forecasting periods is required to establish broader generalizability.