基於深度學習之太陽光電系統對電力品質衝擊預測與分析

Abstract

本研究將對於臺灣中部地區興建併網型太陽光電系統，對於周遭電力系統進行衝擊預測。基於深度學習技術，提出太陽能日照度的預測方法，利用歷史數據建立機器學習預測模型，在本研究中，透過實際太陽光電系統長期收集太陽輻射量與相關環境資料，搭配中央氣象局所提供的衛星雲圖萃取雲量特徵作為訓練資料，使用兩種機器學習方法中常用的遞迴神經網路-長短期記憶網路(Long short-term memory, LSTM)與其變形版本(Gated recurrent unit, GRU)，分別進行短期(未來10分鐘)與中期(未來30分鐘)的太陽輻射量預測。此外，本研究亦使用即時數位模擬系統(Real Time Digital Simulator, RTDS)建立併網型太陽光電系統模型，以長期太陽光電系統資料與預測結果輸入至模型中產生系統衝擊資料。結果顯示，在太陽輻射量預測階段，其中LSTM模型的十分鐘短期預測誤差(Root mean square error, RMSE)為19 W/m2 ，三十分鐘短期預測RMSE為32 W/m2，而GRU模型的十分鐘短期預測RMSE為19 W/m2，三十分鐘短期則為31 W/m2。基於預測結果的一定準度上，藉由預測太陽能輻射量透過電力電子即時數位模擬器進行更為精確的電力系統衝擊分析，此結果有助於提供臺灣電力公司人員調整調度電力決策，提高電網穩健性。

This study establishes a grid-connected photovoltaic (PV) system with a real time digital simulator (RTDS) model based on a real-world power grid in central Taiwan to simulate and analyze the impact of the PV system on the whole power system. Based on deep learning technology, a solar irradiance prediction method is proposed, and the historical data are used to establish the deep learning prediction model. Long-term solar irradiance and related environmental information are collected by the PV system, and additional cloud features are extracted from satellite images to improve the prediction accuracy. With the two common recursive neural network models, the long short-term memory (LSTM) and its modified version (GRU), the irradiance is predicted in a short-term manner (the next 10 minutes) and a medium-term manner (the next 30 minutes). The 10-minute short-term prediction error (RMSE) for the LSTM model is 19 W / m2, while the 30-minute short-term prediction RMSE is 32 W / m2. In addition, the 10-minute short-term prediction RMSE for the GRU model is 19 W / m2, while the 30-minute short-term prediction RMSE is 31 W / m2. Based on these prediction accuracy results, a real-time digital simulator is used to examine the impact of the PV system on the whole power system. The simulation results help Taiwan Power Company to make proper power dispatch decisions, so that the grid stability can be largely improved.