Pengembangan Model Hybrid Spatial Time Series - Machine Learning Untuk Peramalan Permintaan Pelanggan

Prastuti, Mike (2026) Pengembangan Model Hybrid Spatial Time Series - Machine Learning Untuk Peramalan Permintaan Pelanggan. Doctoral thesis, Institut Teknologi Sepuluh Nopember.

Text 7010221013-Doctoral.pdf - Accepted Version Restricted to Repository staff only Download (5MB) | Request a copy

Abstract

Peramalan permintaan memiliki peran penting dalam supply chain management karena menjadi dasar dalam perencanaan produksi, distribusi, pengelolaan persediaan, dan pengendalian biaya. Pada industri semen, permintaan dipengaruhi oleh pola musiman, variasi kalender, keterkaitan antarwilayah, serta kejadian eksternal seperti pandemic COVID-19. Ketidakakuratan peramalan dapat menyebabkan ketidaksesuaian antara jumlah rilis dan kebutuhan aktual, sehingga berpotensi menimbulkan kelebihan persediaan, kekurangan pasokan, peningkatan biaya, maupun penurunan service level. Penelitian ini mengembangkan tiga model hybrid berbasis spatial time series dan machine learning, yaitu GSTARX-Elman RNN (GSTARX-ERNN), GSTARX-Jordan RNN (GSTARX-JRNN), dan GSTARX-Support Vector Regression (GSTARX-SVR), untuk memodelkan permintaan semen multiregional pada empat wilayah pemasaran, yaitu Jawa Timur, Jawa Tengah, Yogyakarta, dan Bali. Model peramalan yang dikembangkan memasukkan variabel eksogen berupa tren waktu, pola musiman bulanan, efek hari raya Idul Fitri, dan intervensi pandemi COVID-19. Hasil penelitian menunjukkan bahwa ketiga model hybrid menghasilkan akurasi peramalan yang lebih baik dibandingkan model nonhybrid, seperti ARIMA, VAR, TSR, ARIMAX, VARX, GSTAR, dan GSTARX. Di antara ketiga model hybrid, GSTARX-ERNN menghasilkan rata-rata MAPE outsample terendah, diikuti oleh GSTARX-SVR dan GSTARX-JRNN. Hasil peramalan selanjutnya digunakan sebagai input optimasi perencanaan agregat untuk menentukan keputusan rilis semen, pengelolaan persediaan, total biaya, dan service level. Hasil optimasi menunjukkan bahwa perbedaan akurasi peramalan berdampak langsung terhadap keputusan rilis semen, akumulasi persediaan, dan total biaya. Model GSTARX-ERNN menghasilkan perencanaan yang lebih efisien dengan total biaya terendah dibandingkan kedua model hybrid lainnya, karena menghasilkan jumlah rilis yang lebih stabil dan persediaan yang lebih terkendali. Hasil analisis sensitivitas menunjukkan adanya trade-off antara efisiensi biaya dan pemenuhan permintaan, di mana peramalan yang terlalu rendah dapat menurunkan service level, sedangkan peramalan yang lebih tinggi cenderung menjaga ketersediaan semen tetapi meningkatkan total biaya. Secara keseluruhan, hasil penelitian menunjukkan bahwa model hybrid spatial time series-machine learning dapat meningkatkan akurasi peramalan permintaan semen serta mendukung pengambilan keputusan terkait rilis, persediaan, dan biaya rantai pasok dan service level secara lebih efisien.
================================================================================================================================
Demand forecasting plays an important role in supply chain management because it serves as the basis for production planning, distribution, inventory management, and cost control. In the cement industry, demand is influenced by seasonal patterns, calendar variations, interregional linkages, and external events such as the COVID-19 pandemic. Forecasting inaccuracy may lead to a mismatch between cement release volumes and actual demand, potentially resulting in excess inventory, supply shortages, increased costs, and a decline in service level. This study develops three hybrid models based on spatial time series and machine learning, namely GSTARX-Elman RNN (GSTARX-ERNN), GSTARX-Jordan RNN (GSTARX-JRNN), and GSTARX-Support Vector Regression (GSTARX-SVR), to model multiregional cement demand across four marketing regions: East Java, Central Java, Yogyakarta, and Bali. The proposed forecasting models incorporate exogenous variables, including time trend, monthly seasonal patterns, the Eid al-Fitr effect, and the COVID-19 pandemic intervention. The results show that the three hybrid models produce better forecasting accuracy than non-hybrid models, such as ARIMA, VAR, TSR, ARIMAX, VARX, GSTAR, and GSTARX. Among the three hybrid models, GSTARX-ERNN produces the lowest average out-sample MAPE, followed by GSTARX-SVR and GSTARX-JRNN. The forecasting results are then used as input for aggregate planning optimization to determine cement release decisions, inventory management, total cost, and service level. The optimization results indicate that differences in forecasting accuracy have a direct impact on cement release decisions, inventory accumulation, and total cost. The GSTARXERNN model produces a more efficient planning outcome with the lowest total cost compared with the other two hybrid models, as it generates more stable cement release volumes and more controlled inventory levels. The sensitivity analysis reveals a trade-off between cost efficiency and demand fulfillment, where underforecasting may reduce the service level, while overforecasting tends to maintain cement availability but increases total cost. Overall, the findings indicate that hybrid spatial time series-machine learning models can improve the accuracy of cement demand forecasting and support more efficient decision-making related to cement release, inventory, supply chain costs, and service level.

Item Type:	Thesis (Doctoral)
Uncontrolled Keywords:	GSTARX, hybrid, machine learning, optimasi biaya, service level, spatial time series, dan supply chain management. GSTARX, hybrid, machine learning, cost optimization, service level, spatial time series, supply chain management.
Subjects:	T Technology > T Technology (General) > T57.62 Simulation T Technology > T Technology (General) > T58.5 Information technology. IT--Auditing T Technology > T Technology (General) > T59.7 Human-machine systems.
Divisions:	Faculty of Intelligent Electrical and Informatics Technology (ELECTICS) > Information System > 55003-(S3) PhD Thesis
Depositing User:	Mike Prastuti
Date Deposited:	16 Jul 2026 03:46
Last Modified:	16 Jul 2026 03:46
URI:	http://repository.its.ac.id/id/eprint/135177

Actions (login required)

View Item