- 學位論文
應用同時擾動隨機近似演算法預測風速及風力發電量
Short-term Wind Speed and Wind Power Forecasting Using Simultaneous Perturbation Stochastic Approximation Algorithm
摘要
風力發電在全球的發電市場中，是一個成長相當快速的能源，然而風能具有間歇性的特性，因此加深了在預測上的難度。對於經濟調度及電力供電而言，正確的預測風能的變化是一件很重要的事。 本研究使用澎湖中屯風場之風速及發電量資料，並使用相關分析及t 檢測，針對輸入資料做前置處理，去除相關性最低之參數。所得之參數使用於本文所提之類神經網路方法進行訓練及預測。 類神經網路在預測研究是一個很好的工具，它提供了一個有效的學習。本論文使用結合同時擾動近似演算法至多層前饋網路及遞迴式類神經網路，比較不同的擾動方式(同時擾動、分層擾動)、網路架構(並接式、串接式、分離式)及輸入，預測未來短期一小時之風速及發電量，最後與傳統預測方法(自我相關移動平均法、持續法)比較其差異。 實驗結果顯示，使用結合SPSA 至多層感知機，對權重同時做擾動，並使用串接式網路進行訓練與測試，R2、平均絕對誤差、均方誤差根之表現為最佳；而使用結合SPSA 至多層感知機，對權重同時做擾動，並使用並接式網路進行訓練與測試，疊代次數及疊代時間為最佳。
並列摘要
Wind energy is one of the most rapidly growing sources for electricity generation in the world. However, operation of wind power is a challenge because of its intermittent characteristics. Therefore, accurate wind power forecasting is necessary for the sake of economic operation in power systems. This research takes advantage of the measurements of wind power generation and wind speed at the same site. Correlation and t-test are applied to pre-process the input data and remove the least correlated input data. Next, the artificial neural networks are used to train and forecast. This thesis presents a novel technique for short-term wind power and wind speed forecasting (1 hour ahead) using a Multi-layer Feed-forward Neural Network (MFNN) and a Recurrent Neural Network (RNN). Neural Network (NN) is one of the best tools for forecasting. It provides a powerful learning ability. Simultaneous Perturbation Stochastic Approximation (SPSA) based train NN may use different perturbations, structures (including cascaded, parallel and separated structures) and inputs. Comparative studies between the proposed methods and traditional methods (including Auto-Regressive Moving Average (ARMA), back-propagation, statistical persistence) are shown in the thesis. The R squared, mean absolute error and root mean square error for the cascaded SPSA-based NN that uses simultaneously perturbing all weightings is better than the other methods. The training CPU times and iterations for the parallel SPSA-based NN that uses simultaneous perturbing all weightings is smaller than the others. The simulation results considering a realistic wind farm including 8 wind generators (each 600kW) and wind speed measurement show the applicability of the proposed method.