基于深度森林算法的电力系统短期负荷预测

Short-term Power Load Forecasting Based on Deep Forest Algorithm

为了提高电力系统短期负荷预测的精确度,解决目前基于机器学习算法的负荷预测需要人为凭经验对超参数进行大量设置和调整的问题,该文将深度森林算法引入了电力系统短期负荷预测领域。深度森林算法包含多粒度扫描阶段和级联森林阶段,具有表征学习的能力。与深度神经网络相比,深度森林算法能够进行高效并行训练,无须大量人为设置和调整超参数。该文选取了某地区实际电力负荷值以及气象因素数据,分别利用了前21天和前40天的数据对深度森林算法进行训练,并将其负荷预测结果与智能算法和传统分类算法的负荷预测结果进行了对比分析。试验结果表明深度森林算法具有高效的电力系统短期负荷预测的能力。

Conventional methods and the famous machine learning algorithms for short-term load forecasting have two shortcomings：（ i） forecasting accuracy is low;（ ii） experiences for the configuration of model hyper-parameters are needed. To mitigate the influence of these shortcomings,deep forest algorithm is applied to short-term load forecasting in power system. Deep forest algorithm,which can do representation learning,includes two procedures： multi-grained scanning procedure and cascading forest procedure. Compared with deep neural network,deep forest algorithm can be trained efficiently in parallel with the default settings for the hyper-parameters of deep forest. The data of systemic actual load and meteorological information are utilized to training the model of deep forest for short-term load forecasting. Tw o models of the forecasting are built in this paper,i. e.,models with the data of previous 21-day and previous 40-day. The forecasting performances of deep forest algorithm are compared with that of numerous intelligent algorithms and conventional classification algorithms,and the results show that deep forest algorithm can forecast the short-term load effectively.