應用機器學習於校園建築能耗預測及契約容量最佳化

Abstract

全球氣候變遷日益嚴重，建築能源消耗佔全球總量 40%，機器學習等新技術有助於建築能耗預測與控制。實時短期預測系統對降低能耗至關重要，對建築能耗管理中有降低尖峰負載的重要性，目前臺灣對負載預測與建築淨零碳排設定了明確目標，由於臺灣因應政策須達到 2050 年 100% 新建建築物及超過 85% 既有建築物為近零碳建築的原因，仍需加強相關技術與管理策略。本研究旨在優化機器學習模型的準確性和各項指標，透過特徵工程和機器學習方法，對校園電耗資料進行分析，並考慮不同氣候變遷情境下的建築能耗趨勢，以改善學校政策和建築環境。此外，預期本研究成果將能結合契約容量最佳化和負載預測，以提供用戶適當的契約容量建議，從而實現用電規劃之目的。 研究流程包括資料收集、模型建構和實驗設計，通過 Python 架構預測模型，並利用氣象和建築歷史資料進行資料前處理和特徵選擇，確保模型的準確性和適用性。通過優化機器學習模型，包括調整演算法、訓練資料庫大小和超參數等手段，我們提高了模型的預測準確性和各項指標。在模型結果分析中，證實了長短期記憶(Long Short-Term Memory, LSTM)在準確度上的優勢，在短期預測中的 MAPE 值達到 4.26%。研究進一步考量長期預測的效果，透過 Shapley Additive Explanations(SHAP) value 分析各特徵對模型的影響，發現關鍵因子為一天前趨勢、一周前趨勢、校園日分類、年中之日及太陽輻射強度。 本研究進一步探討所建立之校園建築能耗模型結果對氣候變遷情境與契約容量最佳化分析的應用，在氣候變遷情境下的分析中，在高排放情境 Shared Socioeconomic Pathways(SSP) 5-8.5 能耗的增加最為明顯，2028 - 2100 年校園建築能耗比例增加將增加 18% - 52 %，未來在建築設計和政策制定上應採取更積極的應對措施，包括安裝高效節能設備和推廣智慧建築管理系統；粒子群優化(Particle Swarm Optimization, PSO)契約容量最佳化分析結果顯示降低原本契約容量 650 元/瓩至 504 元/瓩可以節省 18.1% (約 26 萬元)之電費，證實本研究所建立之校園建築能耗模型此方法的實用性和效益。 本研究的貢獻在於提出了一個綜合考慮負載預測、氣候變遷和契約容量最佳化的建築能源管理架構，為學校和相關機構提供了重要的參考依據和決策支持。未來可以進一步探索模型的應用範圍和準確度提升的方法，以應對不斷變化的能源管理挑戰。

The severity of global climate change has highlighted the importance of energy consumption in buildings, which accounts for 40% of the global total. New technologies such as machine learning are crucial for predicting and controlling building energy consumption. Short-term forecasting systems are essential for reducing energy consumption peaks. Taiwan has set clear goals for load forecasting and achieving net-zero carbon emissions in buildings, driven by policies to ensure that by 2050, 100% of new buildings and over 85% of existing ones will be nearly zero-energy buildings. However, further advancements in relevant technologies and management strategies are needed. This study aims to optimize the accuracy and metrics of machine learning models through feature engineering and machine learning methods. It analyzes campus electricity consumption data, considering different climate change scenarios to enhance school policies and building environments. Additionally, the study expects to integrate optimal contract capacity and load forecasting to provide users with suitable contract capacity recommendations, thereby achieving efficient electricity planning. The research process involves data collection, model construction, and experimental design using Python frameworks. It preprocesses and selects features from meteorological and building historical data to ensure model accuracy and applicability. By optimizing machine learning models, including algorithm adjustments, training database sizes, and hyperparameters, the study improves prediction accuracy and various metrics. The analysis confirms the superiority of Long Short-Term Memory (LSTM) networks in accuracy, achieving a Mean Absolute Percentage Error (MAPE) of 4.26% in short-term predictions. Long-term forecasting effectiveness is further examined using Shapley Additive Explanations (SHAP) values, identifying critical factors such as trends from a day ago and a week ado prior, campus day categorization, day of year, and solar radiation intensity. Furthermore, the study explores the application of the campus building energy consumption model in climate change scenarios and optimal contract capacity analysis. In high-emission scenarios Shared Socioeconomic Pathways(SSP) 5-8.5, energy consumption increases significantly, projecting a rise of 18% to 52% in campus building energy consumption from 2028 to 2100. Future efforts in building design and policy-making should adopt proactive measures such as installing energy-efficient equipment and promoting smart building management systems. The Particle Swarm Optimization (PSO) contract capacity optimization analysis demonstrates potential savings of 18.1% (approximately 260,000 NT dollars) in electricity costs by reducing original contract capacity from 650 NT dollars per kW to 504 NT dollars per kW. This underscores the practicality and benefits of the campus building energy consumption model established in this study. Overall, this research contributes a comprehensive framework for building energy management that integrates load forecasting, climate change considerations, and optimal contract capacity. It provides crucial reference and decision support for schools and relevant institutions. Future research directions include expanding the model's application scope and enhancing accuracy to address evolving energy management challenges.