- 學位論文
電力負載預測準確性之改善研究
The study of electricity load forecasting accuracy improvements
摘要
隨著經濟的快速發展，精確的電力負載預測對於諸如能源發電，電力系統運行安全， 負載單位承諾和能源行銷等許多能源應用至關重要。不精確的電力需求預測也會導致營 運者本身運作成本的增加。因此，過多的需求量預測導致不必要的電力儲備增加，反之， 過低的需求量預測導致電力儲備短缺，從其他供電國家廠商購電的邊際成本將非常昂貴。國際(區域)間電力生產協作廠商均需要精確的電力需求預測。然而，電力需求預測因為 包含諸多複雜多變的因素，諸如氣候因素、社會活動、季節變因等等，至今仍是能源領 域中還未能完全掌握。 支援向量回歸是植基於統計學習理論的預測方法，與傳統回歸模式最大的不同是， 它運用預測模式的結構風險最小化的概念，而非傳統回歸模式直接以訓練誤差值的最小平方法求解回歸參數。相關實證研究顯示，支援向量回歸預測模型由於擁有先進的非線性映射能力，在預測精確度上均優於其他模型。同時，相關研究也指出，針對支援向量回歸預測模型的參數，引進演化算法，進行具有效率機制的搜尋、調校、組合，更有利於預測精確度的提升。 本研究提出克服禁忌搜尋演算法及基因演算法缺點的新式演算法，包括缺乏對禁忌權與搜尋記憶系統的調整(禁忌搜尋法的缺陷)、母群差異性伴隨著搜尋迭代次數的增加而遞減(基因算法的缺陷)等，從而改良支援向量回歸模式預測精確度。本研究應用量子計算中的量子旋轉門(quantum rotation gate)機制除了為禁忌搜尋法重構出更有力的禁忌權架構，改善搜尋記憶系統的效率，即量子禁忌搜尋法；也為基因算法中染色體交配提供更多元的可能性，確保其強大搜尋能力、快速收斂、更少的母群及全域最佳化的能力， 即量子基因演算法。另外，也以混沌理論中的貓映射方程式混合上述量子禁忌搜尋法與量子基因演算法，確保每次迭代計算中母群差異性的存在，避免落入區域最佳解，作為 探索新式混合演化算法改善預測精確度之基礎。 最後以台灣四區域電力需量、年度電力需量及 2014 年全球能源預測競賽(2014 Global Energy Forecasting Competition; GEFCOM 2014)的電力需量資料作為論證，並結合混沌量子禁忌搜尋法與混沌量子基因演算法後的支援向量回歸提高預測精確度的效果。研究結果顯示，結合混沌量子禁忌搜尋法的支援向量回歸模式與結合混沌量子基因演算法的支援向量回歸模式均取得優越的電力需量預測結果，對於電力管理效率提升取得了良好的決策支援數據基礎。
並列摘要
With the rapid economic development, accurate forecasting of electricity load is necessary for many energy applications, such as energy generation, power system operation safety, load unit investment and energy marketing. A forecasting error in electricity load may increase operating cost. Therefore, overestimation of future load may result in oversupply, whereas underestimation of load will result in failure to provide adequate reserve capacity and imply high peaking cost. To generate enough electricity, it depends on every unit of a region to accurately forecast the own need of electricity load. However, forecasting the electricity load can be quite complex because of factors, including climate, social activity and seasonal change. Electricity load forecasting is still not easy. Support vector regression (SVR) is based on the statistical learning theory. Instead of minimizing training errors by using the principle of empirical risk minimization (ERM) of the conventional regression models, SVR minimizes the upper limit of analogical error using the principle of structural risk minimization (SRM) and can learn any training set without error. The empirical result showed that the SVR model could yield higher predictive accuracy than other alternatives because of its advanced non-linear mapping capability. Another known advantage is the use of novel hybrid method of evolutionary algorithm for effective search, adjustment and determination of appropriate parameter combination, which will improve the predictive accuracy by correctly setting the three parameters of the SVR model. In this paper, two novel hybrid algorithms were proposed to overcome the inherent disadvantages of tabu search (TS) and genetic algorithm (GA), including the needs of tabu calibration and elastic memory system to record recent searches (as a shortcoming of the TS algorithm), as well as the lower population diversity and the increase of repeated calculations (as the major disadvantage of GA), so as to improve the predictive accuracy of SVR model. In this study, the quantum rotation gate of quantum mechanics was used to reconstruct a more robust neighborhood structure of a tabu search to improve the efficiency of the elastic memory system, which was referred to as the Quantum Tabu Search (QTS) Algorithm. Meanwhile, this quantum rotation gate was also applied to the chromosomes to provide more possible rotation angles, and thereby ensured its powerful search capability, fast convergence, small population size, and excellent global optimization capacity, which was referred to as the quantum genetic algorithm (QGA). In addition, the cat mapping function was used as a chaotic sequence code generator because this function had better chaotic distribution characteristics to be mixed with QTS and QGA algorithm, respectively, to maintain population diversity, prevent premature convergence and eventually improve forecast accuracy. Finally, three numerical data sets in Taiwan, including electricity load, annual load and competing data (2014 Global Energy Forecasting Competition (GEFCOM 2014)), showed that the proposed models in this study, respectively SVRCQTS and SVRCQGA, were better than alternatives. These superior results provided a robotic decision support for power management.