Laporkan Masalah

Penerapan Metode Neural Hierarchical Interpolation for Time Series (N-HiTS) untuk Long-Horizon Forecasting (Studi Kasus: Data National Energy System Operator)

Talitha Tahara Daffa Mentari, Prof. Dr. Drs. Gunardi, M.Si.

2026 | Skripsi | STATISTIKA

Peramalan permintaan listrik dengan horizon panjang menjadi tantangan karena data per jam memiliki pola pada beberapa skala temporal dan akurasi sulit dipertahankan sering bertambahnya horizon. Penelitian ini menerapkan Neural Hierarchical Interpolation for Time Series (N-HiTS) untuk long-horizon forecasting pada data National Demand (ND) dari National Energy System Operator (NESO), serta membandingkannya dengan Holt-Winters dan Neural Basis Expansion Analysis for Time Series (N-BEATS). Data univariat periode 1 Januari 2023–29 Mei 2026 diagregasi menjadi frekuensi per jam dengan horizon 168 jam. Konfigurasi hyperparameter N-BEATS dan N-HiTS ditentukan melalui grid search berdasarkan MAE validasi. Evaluasi dilakukan dengan skema rolling-origin pada 36 window uji menggunakan Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), dan Mean Absolute Percentage Error (MAPE), lalu dilengkapi uji Friedman dan uji Wilcoxon signed-rank dengan koreksi Holm. N-HiTS memperoleh kinerja terbaik dengan MAE sebesar 1.818,28, MSE sebesar 6,06 × 10?, RMSE sebesar 2.461,41, dan MAPE sebesar 6,67%, mengungguli N-BEATS dengan MAE sebesar 2.033,01 dan Holt-Winters dengan MAE sebesar 795,27, serta unggul pada 25 dari 36 window. Uji Friedman menghasilkan nilai p < 0>

Long-horizon electricity demand forecasting is challenging because hourly demand data exhibit patterns across multiple temporal scales, and accuracy becomes harder to maintain as the horizon grows. This study applies Neural Hierarchical Interpolation for Time Series (N-HiTS) for long-horizon forecasting on National Demand (ND) data from the National Energy System Operator (NESO), comparing it with Holt-Winters and Neural Basis Expansion Analysis for Time Series (N-BEATS). The univariate series, spanning 1 January 2023 to 29 May 2026, was aggregated to an hourly frequency with a 168-hour horizon. Hyperparameters of N-BEATS and N-HiTS were selected via grid search based on validation MAE. Evaluation used a rolling-origin scheme over 36 test windows with Mean Absolute Error (MAE), Mean Squared Error (MSE), Root Mean Squared Error (RMSE), and Mean Absolute Percentage Error (MAPE), complemented by the Friedman test and the Wilcoxon signed-rank test with Holm correction. N-HiTS achieved the best performance, with an MAE of 1,818.28, an MSE of 6.06 × 10?, an RMSE of 2,461.41, and a MAPE of 6.67%, outperforming N-BEATS with an MAE of 2,033.01 and Holt-Winters with an MAE of 2,795.27, and achieving the best performance in 25 of the 36 windows. The Friedman test yielded a p-value below 0.0001, and all model pairs differed significantly at the 5% level. With far fewer parameters and the fastest inference time, N-HiTS offered the best balance between accuracy and computational efficiency.

Kata Kunci : Time Series, N-HiTS, Long-Horizon Forecasting, Electricity Demand, Deep Learning

  1. S1-2026-505391-abstract.pdf  
  2. S1-2026-505391-bibliography.pdf  
  3. S1-2026-505391-tableofcontent.pdf  
  4. S1-2026-505391-title.pdf