基于多相关日场景生成的电动汽车充电负荷区间预测

Interval Prediction of Electric Vehicle Charging Load Based on Scene Generation With Multiple Correlation Days

随着电动汽车(electric vehicle,EV)的持续推广,其强随机性的充电负荷为配电网运行带来挑战。为提高配电网运行可靠性与经济性,提出一种基于多相关日场景生成的EV充电负荷区间预测方法。首先,利用斯皮尔曼秩相关系数,分析待预测日EV充电负荷与其历史日EV充电负荷之间的相关性,找出与待预测日有较强相关性的历史日,构造描述EV充电行为的原始多相关日充电场景集(original multi-correlation day charging scenario set,OMCDCSS)。然后,基于β-变分自编码器(beta-variational auto-encoder,β-VAE),获得与OMCDCSS的概率分布相似且存在时序分布不同的海量生成多相关日充电场景(generating multi-correlation day charging scenario,GMCDCS)。最后,在生成场景集中筛选与已知历史日EV充电负荷数据高度相关的场景,构建相关场景集。基于相关场景集最后一日的数据均值及数据区间分别获得待预测日EV充电负荷确定性预测结果及区间预测结果。对比实验证明,该方法预测区间更可靠,区间宽度更窄。

With the continuous promotion of electric vehicles(EV),its strong random charging load poses challenges to the operation of the distribution network.To improve the reliability and economy of the distribution network operation,an interval prediction method for the EV charging load based on scene generation with multiple correlation days was proposed.First,the Spearman rank correlation coefficient was used to analyze the correlation between the EV charging load on the day to be forecasted and those on the historical days being examined to find the historical days that have strong correlations with the day to be forecasted.The original multiple correlation day charging scene set(OMCDCSS)was constructed to describe the charging behaviors of EVs.Then,large numbers generating multiple correlation day charging scenes(GMCDCS)with similar probability distributions and different timing distributions to the original multiple correlation day charging scenes set were obtained based on the beta-variational auto-encoder(β-VAE).Finally,appropriate scenes with high relevance to the EV charging load data of the known historical days were selected from the generated set to construct a relevant scene set.Based on the data average and data interval of the last day of the relevant scene set,the deterministic and interval prediction results of the EV charging load for the day to be forecasted were obtained.The contrast experiment proves that the proposed method provides a more reliable prediction interval and narrower prediction interval width.