混合電力系統之預測模型及移動窗格最佳化設計

Abstract

本論文延續在混合電力系統中加入預測模型的研究，發展移動窗格(moving window)最佳化設計方法，以不同區間和不同的置換成本進一步分析系統成本變化。此外，我們提出再生能源模組化的概念，透過預測下一個區間的能源需求以及太陽能輻射量，調整混合電力系統的最佳化設計，得出下一個區間系統的能源管理策略和最佳的元件配置，在維持系統供電穩定的前提下降低供電成本。 首先，我們使用混合電力系統數值模型來模擬系統，比起過去所開發之混合電力系統模型，此數值模型在維持系統響應的準確度下，大幅縮短模擬時間並加快系統開發時程。接下來，我們比較了多種機器演算法，最後將較適合的輕量化梯度提升機(Light Gradient Boosting Machine)和極端梯度提升演算法(eXtreme Gradient Boosting, XGBoost)加入系統中，分別做太陽能輻射量和用電負載的預測，之後利用預測模型及能源管理策略，決定氫燃料電池最適當的開啟時機，以此降低氫氣的使用量，達到減少系統供電成本的目的。 其次，本論文發展基於預測模型之移動窗格最佳化設計方法，以兩年的資料來進行案例分析，第一部分設置最長6個月到最短5天等九個不同區間長度，進行移動窗格最佳化設計，結果顯示以一周為區間之設計其供電成本最低，第二部分，則以1%到5%五種不同置換成本，去探討成本與元件置換次數變化，由結果可知置換成本越高，會減少更換元件次數來降低成本。接下來，結合第一與第二部分，討論在不同置換成本下，成本最低且供電穩定之最佳區間長度，結果顯示不論置換成本為多少，以1週為區間之設計，皆可使系統成本降至最低，且同時保持系統穩定。 另外，我們也探討了存在預測誤差的兩種移動窗格設計方法，與準確性為百分之百的完美預測設計，三者之間成本與氫氣消耗量的變化，結果顯示當預測誤差小於5%時，系統供電成本幾乎相同。接著提出再生能源模組化的概念，並開發一台實體再生能源模組，並透過實驗驗證，可透過熱插拔方式來連接模組與混合電力系統，以調節電容量與再生能源發電量。 最後，我們利用混合電力系統實驗，驗證移動窗格最佳化之效益及再生能源模組，此實驗分為未經過移動窗格設計之系統模型和加入該方法設計之模型，將兩種能源管理策略和最佳元件配置結合MATLAB Simulink，建立一套混合電力系統即時響應模型，根據所制定的能源管理策略傳輸指令至實體氫燃料電池，使燃料電池對系統供應電力，透過實際硬體與即時響應模型的整合，驗證移動窗格最佳化設計方法之有效性。另外將模型所模擬出來的二次電池電流響應，經調整後轉成電流指令輸入至電子負載機與電源供應器，觀察系統在接上模組的實際電流響應，再次證明了移動窗格最佳化設計與再生模組應用的可行性。

This thesis proposes a moving-window optimization method for hybrid power systems employing solar panels, batteries, a Proton Exchange Membrane Fuel Cell (PEMFC) and a chemical hydrogen generation system. The moving-window optimization method can adjust system components based on solar and load predictions. Therefore, we could predict the load profiles and solar radiation in the next window to optimize the system design accordingly. The hydrogen consumption and system cost could be further reduced compared with previous optimization methods. First, we simplified the hybrid power system as a mathematical model, which could significantly reduce the simulation time and speed up the system development process while maintaining the accuracy of the system responses. We then selected several machine-learning algorithms to develop prediction models. The Light Gradient Boosting Machine (Light GBM) model and the eXtreme Gradient Boosting (XGBoost) achieved the highest accuracies for forecasting solar radiation and load responses, respectively. Therefore, we integrated these two prediction models with the hybrid power model. We also designed a power management strategy based on the predictions to reduce hydrogen consumption and system costs while maintaining system sustainability. Second, we proposed moving-window optimization for the hybrid power system employing model prediction. We applied a two-year dataset for two analyses. The first analysis is to investigate the impacts of window sizes. We set nine different window sizes, ranging from 6 months to 5 days, where the one-week window outperformed others. The second analysis examined the impacts of replacement cost, ranging from 1% to 5%. The results showed that the system tended to perform less replacement when the replacement costs were high. Finally, we combined these two analyses and concluded that the one-week window size with a 1% replacement cost could significantly reduce the system cost by 10%, compared with the optimal system design without model prediction. We also applied the perfect prediction, i.e., using the actual data instead of the prediction. The results showed that the system costs were almost the same when the prediction errors were less than 5%. Third, we proposed the renewable power module, consisting of batteries and solar panels on a movable platform. The model could be connected to the hybrid power system to adjust the component sizes when necessary. We constructed a physical module and demonstrated that the module could be connected and disconnected to the hybrid power system using hot-swapping. Finally, we conducted experiments to demonstrate the effectiveness of the hybrid power system employing moving-window optimization and renewable power modules. The experiments comprised two parts: the simulation part and the physical part. The simulation part applied MATLAB Simscape ElectricalTM to build a real-time model that sent power management commands to operate the PEMFC, a load meter, and a power supply in the physical part. The results demonstrated the feasibility of the hybrid power system employing moving-window optimization and renewable energy modules.