隨著環保意識提升，太陽能源的發展日漸興起，臺灣整體年均日照長，發展條件良好，並配合智慧電網的應用，能夠把間歇性的太陽能源轉換成穩定可隨時調度的電力，使電力系統變得更加彈性。 由於，太陽能光電系統的產電量主要受到太陽輻射量的影響，因此，針對太陽輻射量的預測進行深入的研究，不同於傳統統計學，本論文採用深度學習的方式，使用卷積神經網路與長短期記憶模型，根據歷史輻射量資訊，對未來逐小時平均太陽輻射強度做預測，利用卷積神經網路提取特徵的特性，以及長短期記憶模型適合預測時間序列資料的特性，提出此兩種架構的混和模型。 本論文致力於長時間的預測，包括預測1日、3日與7日後之太陽輻射量，此外，結合天氣資訊，使預測結果更加準確，然而，考慮到需預測未來的日輻射量，會有缺乏實際觀測天氣數據的狀況，而採用天氣預報的資訊作為判斷標準，預報天氣資訊亦能夠幫助模型預測，以上實驗驗證於臺灣北部與南部地區的案場，證實提出的方法能夠適用於臺灣不同地區的氣候，並且有良好的效果。;With the increasing awareness of environmental protection, the development of solar power has become more popular. In Taiwan, there has been a great potential to develop solar power due to high annual sunlight. With the application of smart grid, we can build a more flexible power system by converting the intermittent solar energy into a more stable and ready to use energy. Because solar radiation is the main factor effecting the power generation of photovoltaic system, therefore the research on the prediction of solar radiation is necessity. Different from the traditional statistic methods, this paper is using a deep learning method in conducting the research. We propose a model using convolutional neural network and long short-term memory, based on the historical solar radiation data to predict the hourly average solar irradiance in the future. This hybrid method is a result of using the unique features of convolutional neural network and the suitable long short-term memory prediction on time series of solar data. This paper is dedicated to the long-term prediction of solar radiation, including the prediction in one, three and seven days ahead. Besides, the prediction accuracy of our model is increased by combining the meteorological information. Considering that there will be a lack of actual weather information, we have shown that the use of weather forecast information is still helpful for prediction modelling. The above experiments were done at various locations in the north and south of Taiwan. Our model has yielded a good result and is suitable for different weather condition in Taiwan.
