Optimasi Hyperparameter Model N-BEATS Menggunakan Optuna untuk Prediksi Harga Telur Ayam Ras di Provinsi Jawa Timur

Abstract

Provinsi Jawa Timur merupakan salah satu sentra produksi telur ayam ras terbesar sehingga dinamika harga telur ayam ras Jawa Timur berpengaruh terhadap laju inflasi. Oleh karena itu, perlu prediksi yang akurat. Namun data harga telur ayam ras memiliki pola nonlinear, fluktuatif, dan nonstasioner, sehingga memerlukan pendekatan deep learning yang adaptif. Penelitian ini menerapkan model neural basis expansion analysis for time series (N-BEATS) yang dioptimasi menggunakan Optuna dengan algoritma tree-structured parzen estimator (TPE), serta mengembangkan ensemble N-BEATS berbasis random initialization untuk meningkatkan performa dan stabilitas prediksi. Periode data yang digunakan dimulai dari 3 Juli 2017 sampai 30 Januari 2026. Penelitian ini bertujuan menentukan konfigurasi hyperparameter terbaik model N-BEATS, mengidentifikasi ukuran ensemble optimal, membandingkan performa N-BEATS dan ensemble N-BEATS, serta memprediksi harga telur ayam ras untuk 130 periode ke depan. Proses optimasi menghasilkan konfigurasi hyperparameter terbaik berupa standard scaler, input size tiga kali horizon, tiga harmonik musiman, tiga derajat basis tren, basis tren Chebyshev, satu block musiman, satu block tren, tiga block identitas, empat lapisan per block, lebar neuron MLP 1024, dan learning rate 0,0022. Hasil pengujian menunjukkan bahwa model ensemble berukuran 18 memberikan performa terbaik dengan persentase penurunan MAE, MAPE, dan RMSE sebesar 10,20%; 11,11%; dan 7,58% dibandingkan model N-BEATS. Prediksi 130 periode ke depan menunjukkan bahwa N-BEATS lebih sensitif terhadap fluktuasi harian, sedangkan ensemble N-BEATS lebih stabil dalam menangkap tren umum, termasuk kenaikan awal yang sejalan dengan Ramadhan dan Idul Fitri serta penurunan pasca Idul Fitri, meskipun lonjakan akibat faktor eksternal belum sepenuhnya terakomodasi.

East Java Province is one of the largest production centers for purebred chicken eggs, making egg price dynamics in East Java influential on the inflation rate. Therefore, accurate forecasting is necessary. However, purebred chicken egg price data exhibit nonlinear, fluctuating, and non-stationary patterns, requiring an adaptive deep learning approach. This study applies the neural basis expansion analysis for time series (N-BEATS) model, optimized using Optuna with the tree-structured parzen estimator (TPE) algorithm, and develops an ensemble N-BEATS model based on random initialization to improve prediction performance and stability. The data cover the period from July 3, 2017, to January 30, 2026. This study aims to determine the best hyperparameter configuration, identify the optimal ensemble size, compare the performance of N-BEATS and ensemble N-BEATS, and forecast egg prices for the next 130 periods. The optimization process produced the best hyperparameter configuration, consisting of a standard scaler, an input size three times the forecasting horizon, three seasonal harmonics, three trend basis degrees, a Chebyshev trend basis, one seasonal block, one trend block, three identity blocks, four layers per block, an MLP neuron width of 1024, and a learning rate of 0,0022. The test results show that the ensemble model with a size of 18 provides the best performance, with reductions in MAE, MAPE, and RMSE of 10,20%; 11,11%; and 7,58%, respectively, compared with the single N-BEATS model. The 130-period-ahead forecast indicates that N-BEATS is more sensitive to daily fluctuations, whereas ensemble N-BEATS is more stable in capturing the general trend.