基於類神經網路之太陽能光電系統發電量預測

Abstract

近年來環保意識急速擴張，因此台灣政府提倡要在2025達到非核家園的目標，希望能在2025年太陽能安裝容量達到20 GW。但現今太陽能安裝容量低靡，無非是因為其不穩定的發電輸出所致。太陽能的發電量會因為天氣因素而有劇烈的波動：如多雲、雨天等因素影響太陽輻射量。本研究提出一太陽能光電系統的發電量預測方法，希望能利用此研究來優化電力調度，其利用環境參數與衛星雲圖當作模型的輸入，透過PCA降維技術挑選出重要的參數，把歷經篩選的參數輸入到Artificial Neural Network模型裡進行訓練，做出了長期預測與短期預測的兩種預測照度模型，其中，太陽輻照度每10分鐘預測一次，這樣如此小的預測間隔與其他研究中使用的間隔不同。 預測間隔越長，預測結果就越準確。 然而，如果採用更長的預測間隔，則太陽輻照度變化在長時間中的變化被忽略。 因此，這項研究使用短時間間隔來預測太陽輻照度。我們分成晴天與陰天探討:晴天的預測，長期預測誤差為139.92 W/m2，短期預測為43.02 W/m2。陰天的預測，長期預測誤差為103.28 W/m2，短期預測為36.99 W/m2。並依此照度的預測結果，與歷史發電量(電流、電壓、功率等)做迴歸曲線分析，找出迴歸方程式與相關性分析，藉此預測未來太陽能系統的發電量。

The use of renewable energy has been actively promoted by the government in Taiwan. For example, the goal of “non-nuclear homeland” is expected to be achieved by 2025. However, the solar system installation capacity remains low, due to the unstable power output caused by unpredictable weather conditions, such as cloudy and rainy days. This study proposes a method for predicting power generation using photovoltaic systems. Various environmental parameters and satellite cloud images serve as the inputs for the prediction model. Important parameters are selected by dimensionality reduction techniques, and the filtered parameters are put into an Artificial Neural Network model to train the model. Both a long-term model and a short-term model for solar irradiance prediction are established. The solar irradiance is predicted every 10 min. Such a small prediction interval is different from the interval used in other studies. The longer the prediction interval, the more accurate the prediction results. However, the solar irradiance variation is largely ignored, if a longer prediction interval is adopted. Thus, this study uses a short interval to predict solar irradiance. The two models deal with two weather conditions: sunny and cloudy days. For sunny days, the long-term prediction error is 139.92 W / m2, and the short-term prediction error is 43.02 W / m2. For cloudy days, the long-term prediction error is 103.28 W/m2, and the short-term prediction is 36.99 W/m2. Furthermore, a regression model is used to analyze the relationship between the solar irradiance prediction results and historical power generation data (i.e., current, voltage, power, and solar irradiance), so the future solar power generation can be predicted.