為了即時有效利用太陽能發電並確保電力品質，準確的極短時光伏發電功率預估為學者努力的方向。這篇報告，採用日照度、溫度等氣象數據，搭配分散式太陽能發電系統發電記錄，運用人工智慧之長短時記憶方法進行太陽能發電模型的訓練與測試。本方法能將長時發電記錄與短時發電資訊透過記憶的方式賦予不同的權重係數，特別適合具時序關係之物理量估測。人工智慧訓練與測試結果分別採用多步與漸進式的預估並加以比較，利用歷史資料的多寡，及時預測5到10分鐘的短期發電成果。相關預測方法與結果及訓練決策方法會加以說明，未來實際應用時機也將討論。

To ensure the effective supply of solar energy and its quality, researchers are looking for better methods to improve the very short-term prediction accuracy of a solar energy harvesting system. In this study, weather information such as solar radiation and temperature, together with the experimental results of a distributed solar power harvesting system is used to train and test by an artificial intelligent, AI algorithm called the long short-term memory (LSTM) method. The LSTM model can assign different weighting coefficients to long-term and short-term memory data, and is particularly suitable for time-series data forecasting. The proposed multi-step and progressive LSTM models are able to provide the up-coming 5 to 10 minutes forecasting of the photovoltaic power system. The detail of the method and prediction results are reported, and the potential application of the machine learning algorithm will be discussed.