Analisis Kinerja Temporal Convolutional Network Berdasarkan Kombinasi Panjang Jendela dan Variasi Parameter

This study aims to analyze the performance of a Temporal Convolutional Network (TCN) in predicting the daily prices of cayenne pepper in Bekasi City, based on variations in window length and architectural parameters. The data used consists of daily cayenne pepper prices from December 1, 2020, to May 30, 2025, obtained from the National Strategic Food Price Information Center (PIHPS). Experiments were conducted across 24 combinations, incorporating window lengths of 7 and 30, dilation patterns of [1,2], [1,2,4], and [1,2,4,8], kernel sizes of 2 and 3, and filter counts of 32 and 64. Evaluation was performed using MSE, RMSE, MAE, MAPE, and R². The results indicate that scenario 10—characterized by a window length of 7, a dilation pattern of [1,2,4,8], a kernel size of 2, and 64 filters—yielded an MSE of 7,861,533, an RMSE of 2,803.84, and an R² of 0.9713. Scenario 11, with a configuration of a window length of 7, a dilation pattern of [1,2,4,8], a kernel size of 3, and 32 filters, resulted in an MAE of 1,569.42 and a MAPE of 2.60%. Conversely, scenario 23—with a window length of 30, a dilation pattern of [1,2,4,8], a kernel size of 3, and 32 filters—produced the lowest performance, with an MSE of 17,226,402, an RMSE of 4,150.47, an MAE of 3,154.08, a MAPE of 5.37%, and an R² of 0.9404. These findings demonstrate that TCN performance is influenced by the interaction between window length and architectural parameters; therefore, the selection of a configuration must be tailored to the specific characteristics of the time-series data being used.