基于相似日的光伏组件积灰损失预测

Prediction of dust accumulation loss in photovoltaic modules based on similar days

为研究灰尘对光伏发电性能的影响,通过搭建的实验台采集清洁与污染光伏组串每天的发电数据,同时监测气象数据,分析积灰及天气对光伏组件发电性能的影响。结果表明,冬季PM2.5质量浓度的上升和春季沙尘暴天气的频发使得光伏组件表面灰尘积累较多,累计发电量损失增长较快,而夏季由于降水增加,灰尘难以积聚在光伏组件上,累计发电量损失增长缓慢。此外,利用DTW(dynamic time warping)算法来寻找相似日。首先通过熵值法计算出各气象参数的权重,然后按日期逆序逐个计算出每个历史日各个气象参数对应的DTW值,再乘以其权重并相加得到历史日的综合DTW值。通过比较各历史日的综合DTW值,选出与当前日最接近的气象相似日。在避开极端天气的情况下,选择数据集中的一部分作为验证集,并对寻找相似日的判据进行优化,选取每天09:00—15:00的数据分为3个时间段进行分析,并设定平均太阳辐照度不小于600 W/m2的条件。优化后,预测模型的评价指标决定系数为0.83,均方根误差为0.22,预测效果显著提升。最后利用该算法为光伏电站制定清洗策略,经过累计发电量损失与清洗成本的对比,确定在长期不降雨情况下,电站应每28天进行一次清洗。

To study the effect of dust on performance of photovoltaic power generation,a laboratory bench was built to collect daily power generation data of clean and polluted photovoltaic strings while monitoring meteorological data to analyze the influence of dust accumulation and weather on power generation performance of photovoltaic modules.The results indicate that,the increase in PM2.5 mass concentration in winter and the frequent occurrence of sandstorms in spring lead to a significant accumulation of dust on surface of the photovoltaic modules,resulting in a rapid increase in cumulative power generation losses.However,in summer,due to increased precipitation,dust is difficult to accumulate on photovoltaic modules,resulting in a slow increase in cumulative power generation losses.In addition,the DTW algorithm is employed to find similar days.Firstly,the entropy method is used to calculate the weights of each meteorological parameter.Then,the DTW values corresponding to each meteorological parameter on each historical day are calculated in reverse chronological order,multiplied by their weights,and added together to obtain the comprehensive DTW value for each historical day.By comparing the comprehensive DTW values of each historical day,the meteorological similar day that is closest to the current day is selected.In order to avoid extreme weather conditions,a portion of the dataset is selected as the validation set,and the criteria for finding similar days are optimized.The data from 9:00 to 15:00 each day is divided into three time periods for analysis,and the condition that the average solar irradiance is not less than 600 W/m2 is set.After optimization,the evaluation index determination coefficient of the prediction model is 0.83,and the root mean square error is 0.22,indicating a significant improvement in prediction performance.Finally,the algorithm is used to develop a cleaning strategy for the photovoltaic power plant.After comparing the cumulative power generation loss with the cleaning cost,it is determined that the power plant should be cleaned every 28 days under long-term non rainfall conditions.