融合时间序列趋势的Dual-ESN机组负荷预测模型

Dual-ESN Prediction Model for Unit Load Fused with Time Series Trend

针对传统模型在机组负荷预测中无法充分捕获内部多变量演化模式的问题,提出了一种基于时间序列的趋势和数值信息融合的双重回声状态网络Dual-ESN(dual-echo state network)机组负荷动态预测模型。首先,引入最小二乘法,对相关的多元历史信息按照局部时间跨度进行趋势拟合。进一步,得到有关过程变化的模式序列,并和原本的数值分别被送入两个独立的储备池,以并行的时间维度进行特征学习。其次,将隐层的高维空间状态送入输出层,融合信息,得到所需要的预测结果。最后,基于山西某工厂660 MW机组装置的真实数据集,进行验证。对比已有预测方法,结果表明所提预测模型在多种性能指标上均有提升。

To solve the problem that the traditional models cannot fully capture the evolution pattern of internal multivariable in unit load prediction,a dual-echo state network(Dual-ESN)dynamic prediction model for unit load based on time series trend and numerical information fusion is proposed. First,the least square method is introduced to perform trend fitting on the relevant multivariate historical information according to the local time span. Furthermore,the obtained pattern sequence related to the process change and the original value are sent into two independent storage pools for feature learning in parallel time dimensions. Second,the high-dimensional space state of the hidden layer is sent into the output layer,and the information is fused to obtain the required prediction results. Finally,the prediction model is verified based on a real data set of a 660 MW unit in a factory in Shanxi Province. Compared with the existing prediction methods,results show that the proposed prediction model improves various performance indicators.