Pengembangan Manajemen Energi Berbasis Deep Q-Learning Multi-Objektif untuk Penurunan LCOE pada Sistem Mikrogrid Hibrida PV-BESS-Diesel PLN ULP Leok

Nainggolan, Ribel Dian Suhaemy
Rosyadi, Marwan
Sujiwanto, Atam Rifa’i

Journal of Authentic Research
Vol. 5 No. 3 (2026): August

Penelitian ini mengembangkan Energy Management System (EMS) berbasis Deep Q-Network (DQN) multi-objektif untuk mikrogrid hibrida PV–BESS–Diesel pada sistem kelistrikan terisolasi PLN ULP Leok, Kabupaten Buol, Sulawesi Tengah, dengan tujuan menurunkan Levelized Cost of Energy (LCOE) tanpa mengorbankan keandalan pasokan listrik. Penelitian dilakukan melalui tiga skenario simulasi yang saling dibandingkan, yaitu skenario eksisting yang seluruhnya mengandalkan PLTD, skenario hasil optimasi tekno-ekonomi mikrogrid PV–BESS–Diesel menggunakan HOMER Pro, dan skenario dispatch energi per jam menggunakan agen Double DQN. Persoalan dispatch dirumuskan sebagai Markov Decision Process dengan ruang state berdimensi 23 variabel, 13 aksi dispatch diskret, serta fungsi reward multi-objektif yang memperhitungkan biaya operasi, konsumsi solar, unmet load, pelanggaran batas SOC baterai, dan energi terbuang secara bersamaan. Mekanisme safety dispatch ditambahkan sebagai batasan operasional agar keandalan suplai tetap terjamin terlepas dari kebijakan yang dipelajari agen. Hasil simulasi selama 8.760 jam operasi menunjukkan bahwa EMS berbasis DQN mampu menurunkan LCOE menjadi Rp2.797/kWh, atau sekitar 19,16% lebih rendah dibandingkan baseline (Rp3.460/kWh) dan 3,12% lebih rendah dibandingkan hasil optimasi HOMER Pro (Rp2.887/kWh). Selain itu, strategi DQN juga berhasil menekan konsumsi bahan bakar solar dan emisi CO₂, menurunkan unmet load secara signifikan, serta meningkatkan renewable penetration menjadi 28,3%. Temuan ini mengindikasikan bahwa strategi dispatch berbasis pembelajaran dapat menjadi pelengkap yang relevan bagi perangkat optimasi tekno-ekonomi seperti HOMER Pro dalam pengoperasian mikrogrid isolated, sekaligus memberikan kontribusi praktis bagi peningkatan efisiensi biaya dan keandalan sistem kelistrikan berbasis diesel seperti ULP Leok. This study develops a multi-objective Deep Q-Network (DQN) based Energy Management System (EMS) for a hybrid PV–BESS–Diesel microgrid serving the isolated network of PLN ULP Leok, Buol Regency, Central Sulawesi, with the aim of reducing the Levelized Cost of Energy (LCOE) without compromising supply reliability. The research compares three scenarios: an existing diesel-only baseline, a techno-economic sizing of the PV–BESS–Diesel configuration using HOMER Pro, and an hourly dispatch strategy generated by a Double DQN agent. The dispatch problem is formulated as a Markov Decision Process with a 23-dimensional state space, 13 discrete dispatch actions, and a multi-objective reward function that simultaneously penalizes operating cost, fuel consumption, unmet load, state-of-charge violations, and excess curtailment. A safety dispatch mechanism is embedded as an operational constraint so that supply reliability is guaranteed regardless of the policy learned by the agent. Simulation results over 8,760 operating hours show that the DQN-based EMS reduces LCOE to Rp2,797/kWh, around 19.16% lower than the baseline (Rp3,460/kWh) and 3.12% lower than the HOMER Pro optimization (Rp2,887/kWh), while also lowering diesel consumption and CO₂ emissions, increasing renewable penetration to 28.3%, and substantially reducing unmet load. These findings indicate that learning-based dispatch can meaningfully complement techno-economic sizing tools in isolated microgrid operation, offering both improved economic performance and reliable supply for diesel-dependent systems such as PLN ULP Leok.