基於Transformer之全天空影像進行估計與預測日射量之系統

Abstract

近年來再生能源發展日益興旺，太陽能作為可持續性發展能源。其發電量與日射量成正關，如能建立一穩定且準確的日射量預測，可加強對緊急狀況之應變能力。在眾多類神經網路類型當中循環神經網路(RNN)已發展多年，其中長短期記憶網路(LSTM)更是被大量使用於具時間序列特性之日射量預測。近年來有學者提出新型態類神經網路模型Transformer，雖其最初目的為語言辨識但因與RNN相似之特性也被大量使用於時間序列之預測。過往之日射量研究多以LSTM為主，然而Transformer模型具有不會梯度爆炸且可同時從多個序列獲取資訊等優點，故本論文嘗試提出一基於Transformer網路為架構之日射量預測模型並以多種效能評估指標與LSTM進行比較。此外，從過往研究可知天氣狀況對日射量有顯著之影響，因此本論文輔以隨機森林(random forest)對數據先進行分類以加強訓練精確度。實驗結果顯示Transformer有不亞於LSTM的預測準確率，在某些指標甚至更勝LSTM。

In recent years, the development of renewable energy has become increasingly prosperous, and solar energy is used as a sustainable energy source. Its power generation is directly related to the amount of sun irradiance .The recurrent neural network (RNN) has been developed for many years, and the long-short-term memory network (LSTM) is widely used in the prediction of sunlight with time series characteristics. In recent years, some scholars have proposed a new type of neural network model Transformer. It is widely used in time series prediction due to its similar characteristics to RNN. In the past, LSTM was mainly used in the research of sunshine amount, but the Transformer model has the advantages of no gradient explosion and the ability to obtain information from multiple sequences at the same time. This study attempts to propose a sunshine forecast model based on the Transformer network and compares it with LSTM with various performance evaluation indicators. In addition, it is known from previous studies that weather conditions have a significant impact on the amount of sunshine, so this study supplemented with Random Forest to classify the data first to enhance the accuracy of training. Experiments show that Transformer has a prediction accuracy rate no less than LSTM, and even better in some indicators .