近期已有許多深度學習模型被提出用於多變量時間序列（MTS）預測的研究。 其中，以 Transformer 為基礎的模型因其具備有效捕捉長期依賴關係的特性，展現 出極大的潛力與應用價值。然而，現有模型在應用於太陽能發電預測時，未能充分 考量太陽能發電本身具有的間歇性與高度依賴氣象條件的特性。此外，這些模型通 常未將未來天氣預報資訊納入考量，使預測的精準度受限。此外，現有的時間序列 預測模型普遍呈現「黑盒」特性，缺乏足夠的可解釋性，導致使用者難以理解模型 決策的依據與運作原理。這些問題凸顯出在太陽能發電預測領域，仍缺乏專門針對 該領域需求所設計的模型。 為填補這項研究空白，本論文提出一種創新的太陽能發電預測模型，結合了過 去的發電數據與氣象資料，以及未來的天氣預報資訊。模型設計上透過將過去與未 來資訊依特定時間間隔切割為多個時間段，從不同時間角度有效地預測未來太陽 能發電量。此外，本論文特別設計多發電預測頭模組，透過多發電區段的預測方法， 有效提升預測準確度，同時賦予模型良好的可解釋性，使得使用者能明確理解模型 如何根據不同因素做出預測。 透過實驗驗證，本論文所提出之模型不僅具備高精確性，並因未來天氣預報資 訊的加入及多預測頭模組設計，性能及解釋性明顯優於傳統方法，能夠為能源管理 者提供清晰明確的決策支援，提升其實務應用價值。;In recent years, numerous deep learning models have been proposed for multivariate time series (MTS) forecasting. Among them, Transformer-based models have demonstrated significant potential and practical value due to their ability to effectively capture long-term dependencies. However, existing models applied to solar power forecasting often overlook the inherent intermittency and strong dependency on meteorological conditions specific to solar energy generation. Furthermore, these models typically fail to incorporate future weather forecasts, thereby limiting their prediction accuracy. In addition, current time series forecasting models generally exhibit "black box" characteristics, lacking sufficient interpretability, which makes it difficult for users to understand the basis and rationale behind model decisions. These limitations highlight the absence of models specifically designed to meet the unique demands of solar power forecasting. To address this research gap, this thesis proposes an innovative solar power forecasting model that integrates historical power generation data, past meteorological data, and future weather forecasts. The model is designed to segment both past and future information into multiple time intervals, enabling effective prediction of future solar power generation from different temporal perspectives. In particular, the model introduces a multi-heads predict module that forecasts power generation across multiple time segments, thereby improving prediction accuracy and enhancing model interpretability. This allows users to clearly understand how the model makes predictions based on various contributing factors. Experimental results validate the effectiveness of the proposed model, demonstrating not only high accuracy but also significantly improved performance and interpretability due to the incorporation of future weather forecasts and the multi-head prediction architecture. These advancements provide clear and actionable decision support for energy managers, enhancing the model’s practical applicability.
