近年來，全球能源結構中綠能發電的比例不斷增加，但同時也帶來了能源供應的不穩定性和不可預測性的挑戰。為了解決這些問題，儲能系統成為了重要的解決方案。儲能系統可以暫時將綠能發電過剩的能量儲存起來，在發電量不足或電費較高的時候應急使用，妥善利用儲能系統可以平衡能源供需，維持電網穩定。這樣的措施有助於提高能源利用效率，實現更可持續的綠能發展。而在儲能系統上使用機器學習演算法，觀察和學習家戶用電的習慣，利用時間電價表改變充放電的時機，可進一步節省家戶用電費用的支出。 本論文旨在比較Long Short-Term Memory (LSTM)、Gated Recurrent Unit (GRU)和Simple Recurrent Unit (SRU)這三種循環神經網絡Recurrent Neural Network (RNN)演算法應用在家用儲能系統用電量和發電量的預測準確性和訓練效能。實驗數據顯示，這三種演算法在訓練集和測試集的預測準確率不相上下，而在訓練時間上SRU可較 LSTM 快43%。訓練速度的提升，使SRU更適合處理複雜的應用計算。在採用演算法輔助充放電決策後，並考量了充放電損耗後，家用儲能系統可利用時間電價的優惠，再節省4.5%的電費支出。 ;In recent years, the proportion of green energy power generation in the global energy structure has been increasing, but it has also brought challenges of instability and unpredictability of energy supply. In order to solve these problems, energy storage systems have become an important solution. Energy storage systems can temporarily store excess energy generated by green energy for emergency use when power generation is insufficient or electricity bills are high. Proper use of energy storage systems can balance energy supply and demand and maintain grid stability. Such measures can help improve energy efficiency and achieve more sustainable green energy development. Using machine learning algorithms on energy storage systems to observe and learn household electricity consumption habits, and using time electricity price tables to change the timing of charging and discharging, can further save household electricity costs. This paper aims to compare the electricity consumption of three types of Recurrent Neural Network (RNN) algorithms: Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU) and Simple Recurrent Unit (SRU) when applied to home energy storage systems. and power generation prediction accuracy and training efficiency. Experimental data shows that the prediction accuracy of these three algorithms in the training set and test set is similar, and SRU can be 43% faster than LSTM in training time. The improvement in training speed makes SRU more suitable for processing complex application calculations. When dealing with abnormal data, LSTM performs better in data prediction due to its more complete gating mechanism. After using algorithms to assist charging and discharging decisions, home energy storage systems can save another 4.5% in electricity bills.
