近年來再生能源因技術發展、能源轉型及淨零碳排等政策目標推動下蓬勃發展，尤其以太陽光電之技術相當成熟穩定，也普遍為社會大眾接受，從評估、規劃、建置亦可帶動相關產業發展，完工後的運轉與維護是未來不可或缺的一部分。其中以大型光電案場來說，實務上會使用直流發電比（Array Ratio,RA）來評估整個案場甚至單組串並的性能，若對維護加入預測性，包括預測硬體故障、預測清洗例程等可以減少因長期累積的髒污或衰減導致發電損失，預知上述問題並及時處理即可提高發電品質，如何以深度學習進行RA值預測，發現潛在問題並減少太陽光電案場維運的人力成本，為本文研究之方向。
本文藉由太陽能實務上會遇到的故障進行分析及處理辦法說明，再以長短期記憶神經網路（Long Short-Term Memory, LSTM）以及門控循環單元網路（Gated Recurrent Unit, GRU）透過平均平方誤差（Mean Square Error, MSE）分析預測RA值後計算直流電流I之準確性，分別透過m,n時間序列進行預測，預測的目標是建立在歷史數據m的基礎上，使用模型預測未來時間n，藉由不斷調整m和n的數值，並觀察各預測模型適用不同的需求情況。
實驗結果證明，若要執行太陽光電系統RA值的n=1小時至24小時短期預測中以m=24小時的數據LSTM預測的結果最佳，若需要進行n=7天至30天之長期預測時選擇m=30天的數據並執行GRU預測較為適合，最後以直流電流進行短期預測所得出的MSE誤差也較直接預測RA值的低，所以選擇最適合及經濟的類神經網路預測是相當重要的。

In recent years, renewable energy has grown significantly due to technology advancements, energy transition, and net-zero energy policies. Photovoltaic technology, among the renewable sources, is now mature and widely accepted, evaluating, planning, and implementing solar PV projects can also drive related industries' development, operation and maintenance are indispensable components of the post-construction phase. The Array Ratio (RA) is used to assess large-scale photovoltaic projects. By using predictive maintenance, like forecasting hardware failures and cleaning routines, power generation losses from debris accumulation or degradation can be reduced. This paper explores how Deep Learning can predict RA to identify potential problems and lower labor costs for operating and maintaining photovoltaic projects.
In this paper, explains how to analyze and deal with faults that occur in practical solar energy systems. Long Short-Term Memory (LSTM) and Gated Recurrent Unit (GRU) neural networks are used to predict the DC current (I) accuracy by predicting RA using Mean Square Error (MSE) analysis. The predictions are performed on time series data with different values of m and n. The goal of the predictions is to build a model based on historical data (m) and use it to predict future time periods (n). By continuously adjusting the values of m and n and observing the performance of different prediction models, the study aims to identify the most appropriate models for different requirements and scenarios.
Experimental results have shown that for short-term forecasting of the RA in solar photovoltaic systems, with n ranging from 1 hour to 24 hours, using m equal to 24 hours of data for LSTM prediction yields the best results. For long-term forecasting with n ranging from 7 days to 30 days, selecting m equal to 30 days of data and using GRU for prediction is more suitable. Furthermore, when using direct current (I) for short-term prediction, the MSE is lower than the RA. Therefore, choosing the most suitable and cost-effective neural network model for prediction is of paramount importance.

[1] 台灣電力公司，台電系統歷年裝置容量，[Online]. Available:https://www.taipower.com.tw/tc/page.aspx?mid=212&cid=119&cchk=792b6538-55d3-4ded-a2fc-a9eef608febb
[2] Y. Zhao, J. Palma, J. Mosesian, R. Lyons, and B. Lehman, "Line–Line Fault Analysis and Protection Challenges in Solar Photovoltaic Arrays", IEEE TIE, pp. 3784–3795, JUN 2012.
[3] Y. Zhao, R. Ball, J. Mosesian, J. Palma, and B. Lehman, "Graph-Based Semi-supervised Learning for Fault Detection and Classification in Solar Photovoltaic Arrays", IEEE Trans. Ind. Electron., vol. 9, no. 5 , pp. 2848–2858, MAY 2015.
[4] Y. Zhao, L. Yang, B. Lehman, J. Palma, J. Mosesian, and R. Lyons, "Decision tree-based fault detection and classification in solar photovoltaic arrays", IEEE APEC, pp. 93–99, MAR 2012.
[5] Y. Zhao, B. Lehman, R. Ball, J. Mosesian, and J. Palma, "Outlier Detection Rules for Fault Detection in Solar Photovoltaic Arrays", IEEE APEC, pp. 2913–2920, MAR 2013.
[6] S. Shamshirband, T. Rabczuk, and K. W. Chau, "A Survey of Deep Learning Techniques: Application in Wind and Solar Energy Resources", IEEE Access, vol. 7 , pp. 164650–164666, NOV 2019.
[7] C. H. Liu, J. C. Gu, and M. T. Yang, "A Simplified LSTM Neural Networks for One Day-Ahead Solar Power Forecasting", IEEE Access, vol. 30, no. 5 , pp. 17174–17195, JAN 2021.
[8] S. Al-Dahidi, O. Ayadi, M. Alrbai, and J. Adeeb, "Ensemble Approach of Optimized Artificial Neural Networks for Solar Photovoltaic Power Prediction", IEEE Access, vol. 7, pp. 81741–81758, JUN 2019.
[9] R. Jan, P. Chaiwiwatworakul, and N. Chawphongphang, "Forecasting the Global and Diffuse Solar Irradiance of Bangkok's Tropical Climate using Long-Short Term Memory (LSTM) Technique", IEEE ICBIR, pp. 1297–1302, JUN 2023.
[10] A. Prajesh, P. Jain, and Md. K. Anwar, "Solar Irradiance Forecasting Using Ensemble Models of Machine Learning", IEEE IAS, pp. 1–6, MAR 2023.
[11] M. A. Jallal, A. Yassini, S. Chabaa, A. Zeroual, and S. Ibnyaich, "Ensemble Learning Algorithm-based Artificial Neural Network for Predicting Solar Radiation Data", IEEE DASA, pp. 526–531, NOV 2020.
[12] N. Kim, and B. Parkhideh, "Comparative Analysis of Non-Isolated and Isolated Type Partial-Power Optimizers for PV-Battery Series Inverter Architecture", IEEE ECCE, pp. 6207–6213, DEC 2018.
[13] N. Muller, J. Barria, and L. Reyes–Chamorro, "Flexible Power Point Tracking for Active Power Regulation in String-Inverter PV Plants with Power Optimizers", IEEE eGRID, NOV 2021.
[14] M. A. Yagoub; A. A. Elserougi, and T. H. Abdelhamid, "Assessment of PV String Optimizer Circuit Versus String and Central Inverters During Partial Shading Conditions", IEEE MEPCON, pp. 14–19, DEC 2021.