Resilient Data-Driven Optimization of Multi-Site Green Power Wheeling and Storage Dispatch

A mixed-integer optimization model that coordinates cross-site green-power wheeling, rooftop solar, and battery storage across a corporate group's offices, malls, and warehouses — minimising electricity cost and demand-charge risk while maximising renewable self-use. Validated on a full year of 15-minute data from three real sites in northern Taiwan.

Meeting corporate net-zero targets — RE100, ESG disclosure, green-supply-chain rules — increasingly hinges on how a company sources and routes renewable electricity, not just how much it buys. In Taiwan’s green-power market, “wheeling” (轉供) allocates renewable attributes by contract rather than physical flow, which leaves companies with two persistent problems: generation and demand rarely line up in time, and multiple sites with different load profiles are hard to coordinate — so one site curtails surplus solar while another buys extra certificates to cover a shortfall.

Run with industry partner NextDrive Inc. (聯齊科技) under the NSTC Industry–Academia program, this project builds a data-driven, resilient optimization model that dispatches green-power wheeling, rooftop solar, and battery storage across a company’s sites as a single system — from the electricity buyer’s perspective rather than a single building’s.

The model. The problem is formulated as Mixed-Integer Linear Programming (Python · CVXPY · HiGHS solver). Decision variables span grid (brown) and purchased-green power, cross-site solar wheeling (BGES), PV self-use and battery charging, storage charge/discharge with a binary no-simultaneous guard and state-of-charge tracking, and a two-tier over-contract-capacity penalty. The objective minimises total operating cost — energy purchases plus tiered demand penalties — with a deliberately tiny wheeling-cost coefficient that nudges the model to share power internally before buying from the grid. Constraints enforce power balance, renewable conservation with self-use priority, storage limits, a minimum renewable-share (RE%) threshold per site, and a no-self-wheeling rule.

The data. The model was validated on a full year (May 2024 – April 2025) of 15-minute load and solar data — 35,040 timesteps — from three heterogeneous sites of a northern-Taiwan corporate group: a daytime office (184 kW contract, rooftop PV + storage), an all-day high-load mall (1,730 kW, the main green-power consumer), and a warehouse whose 3,400 kWp of solar far exceeds its own flat load, letting it act as a virtual power plant and the group’s main supplier.

Backtested against a no-dispatch baseline, the optimized model cut grid reliance, held every site under its contracted capacity all year, and lowered the group’s total bill while accumulating renewable-energy certificates — with results framed as deployable recommendations for integration into NextDrive’s energy-management platform.

Key findings

After optimization, grid demand at all three sites stayed strictly below contracted capacity (184 / 1,730 / 1,201 kW) for every one of the year's 35,040 fifteen-minute intervals — eliminating demand-penalty risk. Brown-electricity demand fell significantly at each site (permutation test p ≈ 0; −44%, −28%, −27% of baseline mean).

Surplus rooftop solar from a 3,400 kWp warehouse is wheeled in real time to high-consumption sites — supplying ~3.4% of the mall's annual demand and ~4.6% of the office's — instead of being curtailed. On low-sun days the model automatically suspends wheeling to protect each site's own supply.

Even with purchased green power priced above grid rates (NT$6.7/kWh), the group's total annual electricity bill fell ~NT$2.9M versus baseline — the mall alone saving ~NT$3.1M (−7.6%) — while the group banked 2,831 extra renewable-energy certificates (T-RECs) for Scope-2 and RE100 reporting.

企業要達成淨零目標——RE100、ESG 揭露與綠色供應鏈要求——關鍵日益不只在於購買多少綠電，更在於如何取得與調度再生能源。臺灣綠電市場以「轉供」運作，其本質是對綠電屬性的契約計算認證，而非實體電力流動，使企業長期面臨兩大瓶頸：發電與用電的時序經常不一致，以及負載特性各異的多場域難以協調—— 導致某一場域棄光，另一場域卻須額外購買憑證來補足缺口。

本計畫在國科會產學合作計畫下與業界夥伴聯齊科技（NextDrive）合作，建構一套資料驅動的強韌性最佳化模型，從購電業者而非單一建築的角度，將跨場域綠電轉供、屋頂太陽能與儲能視為單一系統進行調度。

模型方法。 問題以混合整數線性規劃（MILP）建模，於 Python 環境以 CVXPY 搭配 HiGHS 求解器實作。決策變數涵蓋市電（灰電）與外購綠電、跨場域太陽能轉供（BGES）、太陽能自用與充電、儲能充放電（以二元變數避免同時充放電並追蹤 SOC），以及兩級超約容量罰則。目標函數最小化整體營運成本——購電支出加上分級需量罰則 ——並將轉供成本係數設為極小值，誘導模型優先進行場域間內部共享，再向電網購電。限制條件涵蓋電力供需平衡、再生能源守恆與自用優先、儲能容量上下限、各場域最低再生能源占比（RE%）門檻，以及禁止場域自我轉供。

驗證資料。 模型以臺灣北部某企業集團三個異質場域、全年度（2024 年 5 月至 2025 年 4 月）15 分鐘解析度的負載與太陽能資料（共 35,040 筆時序）進行驗證：日間使用型的辦公樓（契約容量 184 kW，配置屋頂太陽能與儲能）、全天候高負載的商場（1,730 kW，主要綠電需求端），以及 3,400 kWp 太陽能遠超自身平穩低負載的倉儲——使其得以扮演虛擬電廠與集團主要供給端的角色。

相較於無調度的基準情境，最佳化模型成功降低市電依賴、全年將各場域用電維持在契約容量之下，並在累積再生能源憑證的同時降低集團總電費——研究成果並轉化為可落地的建議，供整合至聯齊科技的能源管理平台。