Peramalan Penjualan Produk Mie Instan Menggunakan Model Hibrida CNN–BiLSTM dengan Optimasi Genetic Algorithm (Studi Kasus: PT. XYZ Banjarmasin)

Yamani, Muhammad Iqbal Baiduri (2026) Peramalan Penjualan Produk Mie Instan Menggunakan Model Hibrida CNN–BiLSTM dengan Optimasi Genetic Algorithm (Studi Kasus: PT. XYZ Banjarmasin). Other thesis, Institut Teknologi Sepuluh Nopember.

Text 5026221103-Undergraduate_Thesis.pdf Restricted to Repository staff only Download (15MB) | Request a copy

Abstract

Industri Fast Moving Consumer Goods (FMCG) menghadapi tantangan besar dalam memprediksi permintaan distributor akibat pola penjualan yang sangat fluktuatif. Kesalahan peramalan (forecasting) dapat memicu kelebihan stok maupun kekurangan stok yang merugikan perusahaan secara operasional. Tugas akhir ini bertujuan membangun model peramalan penjualan mingguan menggunakan arsitektur hibrida Convolutional Neural Network dan Bidirectional Long Short-Term Memory (CNN-BiLSTM) yang dioptimasi oleh Genetic Algorithm (GA). Data historis penjualan produk mie instan terlaris dari tahun 2019 hingga 2025 dikonversi menggunakan sliding window (lag 8 minggu) menjadi format supervised learning. Evaluasi internal pemodelan dilakukan melalui skema Walk-Forward Cross Validation. Hasil pengujian pada empat skenario pembagian proporsi data (60:40 hingga 90:10) menunjukkan bahwa arsitektur hibrida CNN-BiLSTM justru mengalami kesulitan dalam menangkap volatilitas ekstrem data penjualan business-to-business (B2B), dengan tingkat Mean Absolute Percentage Error (MAPE) terbaik hanya sebesar 25,56% pada proporsi 90:10. Sebaliknya, model BiLSTM tunggal yang dioptimasi oleh GA terbukti memberikan akurasi tertinggi dengan MAPE mencapai 16,72%. Namun demikian, model BiLSTM murni tanpa optimasi GA mencatatkan performa yang sangat kompetitif dengan selisih error yang tipis, yakni MAPE 17,25%. Mengingat optimasi GA menuntut trade-off waktu komputasi yang masif hingga berjam-jam dibandingkan BiLSTM murni yang hanya memakan waktu puluhan detik, BiLSTM murni menjadi opsi yang paling rasional. Kesimpulannya, untuk karakteristik deret waktu univariat penjualan FMCG yang sarat lonjakan acak, penggunaan model BiLSTM murni lebih direkomendasikan dibandingkan arsitektur hibrida kompleks maupun optimasi GA guna mendukung perencanaan persediaan yang akurat dan sangat efisien secara komputasi.
===============================================================================================================================
The Fast Moving Consumer Goods (FMCG) industry faces significant challenges in predicting distributor demand due to highly fluctuating sales patterns. Forecasting errors can trigger overstocking or stockouts, which operationally harm the company. This study aims to build a weekly sales forecasting model using a hybrid Convolutional Neural Network and Bidirectional Long Short-Term Memory (CNN-BiLSTM) architecture optimized by a Genetic Algorithm (GA). Historical sales data of the best-selling instant noodle product from 2019 to 2025 were converted using a sliding window (8-week lag) into a supervised learning format. Internal modeling evaluation was conducted through a Walk-Forward Cross Validation scheme. Test results across four data splitting scenarios (60:40 to 90:10) showed that the hybrid CNN-BiLSTM architecture actually struggled to capture the extreme volatility of business-to-business (B2B) sales data, achieving its best Mean Absolute Percentage Error (MAPE) of 25.56% at the 90:10 proportion. Conversely, the standalone BiLSTM model optimized by GA proved to deliver the highest accuracy with a MAPE reaching 16.72%. However, the pure BiLSTM model without GA optimization recorded highly competitive performance with a narrow error margin, achieving a MAPE of 17.25%. Considering that GA optimization demands a massive computational time trade-off of up to several hours compared to the pure BiLSTM which only takes tens of seconds, the pure BiLSTM emerges as the most rational choice. In conclusion, for the characteristics of univariate FMCG sales time series laden with random spikes, the use of a pure BiLSTM model is highly recommended over a complex hybrid architecture or GA optimization to support accurate and computationally efficient inventory planning.

Item Type:	Thesis (Other)
Uncontrolled Keywords:	CNN-BiLSTM, FMCG, GA, Hibrida, Peramalan, Forecasting, Hybrid
Subjects:	T Technology > T Technology (General) > T174 Technological forecasting
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 57201-(S1) Undergraduate Thesis
Depositing User:	Muhammad Iqbal Baiduri Yamani
Date Deposited:	29 Jul 2026 07:20
Last Modified:	29 Jul 2026 07:20
URI:	http://repository.its.ac.id/id/eprint/138592

Actions (login required)

View Item