基于CNN-LSTM分位数回归的母线负荷日前区间预测

针对部分工业类母线负荷波动较大,传统点预测方法难以准确预测的问题,文中提出一种基于卷积神经网络(CNN)与长短期记忆网络分位数回归(QRLSTM)组合的母线负荷日前区间预测模型。首先,针对工业类负荷功率的高频波动,采用去噪自编码器对历史负荷数据进行降噪处理;然后,利用基于时间分布层封装的一维CNN网络进行负荷特征提取和压缩,以提升整个模型的学习效率;最后,建立含有注意力机制的QRLSTM模型进行特征学习,得到不同分位数下的负荷区间预测结果。对工业类和居民商业类2种典型的220 kV母线负荷进行了负荷日前区间预测测试,并与常规的分位数回归方法进行了对比。结果表明,文中方法获得的预测结果总体上区间覆盖率更大、区间平均宽度和区间累计偏差均更小,预测效果更好。

Day-ahead interval prediction of bus load based on CNN-LSTM quantile regression

It is difficult to predict accurately the bus load by traditional point prediction methods due to the violent fluctuation of some industrial bus load. A day-ahead interval prediction model of bus load based on the combination of convolutional neural network (CNN) and quantile regression long short-term memory (QRLSTM) is proposed in this paper. Firstly, the denoising auto-encoder is used to obtain the historical load data by de-noising the high frequency fluctuation of industrial load power. Then a one-dimensional CNN network encapsulated by time distribution layer is used to extract load features to improve the learning efficiency of the whole model. Finally, the QRLSTM model with attention mechanism is established for feature learning, and the load interval prediction results under different quantiles are calculated. The day-ahead interval prediction results of bus load are obtained by an industrial 220 kV bus and a commercial-residential 220 kV bus. The results show that the proposed prediction method has generally larger interval coverage, smaller interval mean width, smaller interval cumulative deviation and higher prediction effectiveness than the conventional quantile regression method.