| DC 欄位 |
值 |
語言 |
| DC.contributor | 資訊工程學系 | zh_TW |
| DC.creator | 張至妤 | zh_TW |
| DC.creator | Chih-Yu Chang | en_US |
| dc.date.accessioned | 2022-7-18T07:39:07Z | |
| dc.date.available | 2022-7-18T07:39:07Z | |
| dc.date.issued | 2022 | |
| dc.identifier.uri | http://ir.lib.ncu.edu.tw:88/thesis/view_etd.asp?URN=109522055 | |
| dc.contributor.department | 資訊工程學系 | zh_TW |
| DC.description | 國立中央大學 | zh_TW |
| DC.description | National Central University | en_US |
| dc.description.abstract | 隨著再生能源的興起，太陽能源的發展日漸重要，臺灣因年均日照長，發展太陽能條件良好，若能有穩定的產電預測，將使整體電力系統更彈性、更能因應劇烈的太陽變化。
太陽能發電量主要受到日射量多寡來轉換，因此本篇論文將以日射量的預測進行研究。而短期的日射量變化容易受到天空狀況影響，雲的狀態將影響當下太陽照射到地面的輻射量，故本篇論文提出以深度學習的方式，使用卷積神經網路與長短期記憶模型，根據歷史的日射量及全天空影像資訊，來預測未來逐分鐘的日射量。先藉由全天空影像判斷不同的天空狀態，再將日射計及全天空影像儀蒐集的資訊當作訓練特徵並結合，根據不同的天空狀態類別訓練出不同的深度學習模型，預測時即能根據當下天空狀態選擇適合的模型得到未來日射量結果。其中訓練模型時，透過全天空影像藉太陽位置演算法得到影像中太陽位置，再分析太陽附近灰階度特徵，藉由太陽附近灰階度特徵及全天空影像特徵來輔助日射量特徵進行未來日射量預測。
整體實驗藉由分月份的方式進行訓練、驗證及測試，所有實驗結果將在RMSE、RMSPE、MAE、MAPE上做評估，並根據預測的時長分開統計，實驗顯示藉由不同天空狀態分模型的方式，並影像及數值特徵結合的方法，能夠以較低的誤差預測出日射量。 | zh_TW |
| dc.description.abstract | Solar energy is becoming more and more popular and important with the rise of renewable energy. Taiwan experiences long, hot summers, and short, mild winters. This makes Taiwan an ideal location for solar energy development. If there is a stable solar energy power generation, the overall power system will be more flexible and able to respond quickly to high fluctuations in supply and demand.
Solar power generation is mainly converted by the amount of solar irradiance. Short-term changes in solar irradiance are easily affected by sky conditions, and the state of clouds. In this work, a very short-term solar irradiance prediction mechanism based on automatic sky conditions classification is proposed. First, different sky states are judged by the all-sky images. Multiple deep learning models are constructed by different sky states using historical irradiance values and all-sky images as features. The deep learning model, which combines convolutional neural network and long- short-term memory network, can extracted both image and numerical features. Moreover, the solar position in the image is obtained through the solar position algorithm, and then the around pixels’ gray-scale features are analyzed and then input to the training model too.
The overall experiment is trained, validated and tested by monthly. All experimental results will be evaluated on RMSE, RMSPE, MAE, and MAPE. Results have shown that incorporating sky conditions information can capture different characteristics of irradiance variation under different sky states, and the method combined with images and numerical features can predict the irradiance with low error. | en_US |
| DC.subject | 深度學習 | zh_TW |
| DC.subject | 卷積神經網路 | zh_TW |
| DC.subject | 長短期記憶 | zh_TW |
| DC.subject | 太陽輻射量 | zh_TW |
| DC.subject | 全天空影像 | zh_TW |
| DC.subject | 太陽位置演算法 | zh_TW |
| DC.subject | Deep learning | en_US |
| DC.subject | Convolutional neural network | en_US |
| DC.subject | Long short-term memory | en_US |
| DC.subject | Solar Irradiance | en_US |
| DC.subject | All-sky images | en_US |
| DC.subject | Solar Position Algorithm | en_US |
| DC.title | 卷積長短期記憶神經網路結合全天空影像特徵之短期日射量預測模型 | zh_TW |
| dc.language.iso | zh-TW | zh-TW |
| DC.title | Very Short-Term Solar Irradiance forecasting using convolution Long-Short-Term Memory Network with All-Sky images Features | en_US |
| DC.type | 博碩士論文 | zh_TW |
| DC.type | thesis | en_US |
| DC.publisher | National Central University | en_US |