探索資料融合技術應用於多元天氣資料預測太陽能發電之潛力

Abstract

在太陽能全球化發展的趨勢下，太陽能潛勢預測面臨日益差異化的地面天氣資料觀測條件。地面觀測資料為傳統上最準確的天氣資料來源，然而其空間分布不均，使作為預測主要輸入的天氣資料在可靠性上產生疑慮。本研究旨在探討氣象輸入資料的品質如何影響預測結果，進而提升太陽能潛勢預測的整體可靠性。 本研究採用兩種天氣資料來源：中央氣象署地面觀測資料 (點資料) 與 ERA5 重分析資料 (格點資料)，並設計八種天氣輸入組合，涵蓋單一來源、雙資料來源使用策略 (包含平行化輸入、Kriging with External Drift 與 Kriging with Radar-Based Error Correction 兩種資料融合技術) 及關鍵地面變數的納入與否。預測模型採用長短期記憶 (LSTM) 神經網路，並在相同模型架構與訓練流程下進行系統性比較分析。為模擬全球不同地區的觀測資源差異，進一步設計十三種地面觀測密度情境，並評估各輸入組合於不同情境下之預測表現。 研究結果指出，在使用單一資料來源的情況下，地面觀測資料於太陽能潛勢預測的優勢範圍約為氣象站距離小於 75 公里；而當觀測資源稀缺時，格點資料則展現出良好的穩定性與應用潛力。在 50–75 公里這一關鍵距離範圍內，雙資料來源策略的表現優於僅使用格點資料，其中以平行化輸入法效果最佳，其次為 KRE 融合方法。此外，研究亦發現三項地面觀測專屬變數：日照率、日照時數與最大紫外線指數，對於提升預測精度具有正面效益。

Under the global trend of solar energy development, solar power potential prediction faces increasingly diverse ground weather observation conditions. Although ground-based weather data are traditionally the most accurate source, their uneven spatial distribution raises concerns regarding the reliability of weather inputs for prediction. This study aims to examine how the quality of weather inputs affects prediction outcomes, in order to improve the overall reliability of solar power potential prediction. Two sources of weather data were used: point-based observations from the Central Weather Administration (CWA) and gridded ERA5-reanalysis data. Eight input combinations were designed to represent different strategies, including single-source inputs, dual-source approaches (parallel input and two data merging techniques: Kriging with External Drift [KED], and Kriging with Radar-Based Error Correction [KRE]), and inclusion of key ground-based variables. A Long Short-Term Memory (LSTM) neural network was employed as the prediction model, and all combinations were evaluated under the same model architecture and training procedure. To simulate varying observational resource conditions worldwide, 13 weather station density scenarios were further constructed to assess performance under different levels of data availability. The results show that ground-based data outperform gridded data when the distance to the nearest station is less than 75 km. When observational resources are limited, ERA5 gridded data demonstrates better stability and applicability. Within the critical 50--75~km range, dual-source strategies showed better performance than gridded-only inputs, with the parallel input method yielding the best performance, followed by the KRE data merging method. In addition, three CWA-specific variables: sunshine duration, sunshine percentage, and maximum UV index were found to have positive effects on prediction accuracy.