改进多尺度嵌套长短期记忆神经网络的虚拟电厂负荷预测方法

Improved Multiple Scale Nested Long Short-term Memory Neural Network Load Forecasting Method for Virtual Power Plants

虚拟电厂负荷受多种因素的综合影响,表现出高度非线性和动态变化的特性,这些特性导致其难以预测。现有预测方法具有诸多局限性,如对负荷时间序列特征提取不足、对负荷影响因素考虑不全面等。这些都会导致预测精度难以进一步提高。为此,提出一种改进的虚拟电厂负荷预测方法,该方法基于多尺度嵌套长短期记忆(MSNLSTM)神经网络,构建多层次的长短期记忆网络用于提取负荷序列不同时间尺度的模式,以深入学习负荷的内在周期性和相关性特征。同时,引入外部因素数据作为网络的输入,以增强对负荷影响因素的建模能力。实验证明,相对于单一长短期记忆网络和传统预测方法,所提模型能够提高日前和周前负荷预测的精度。

The virtual power plant load is affected by a variety of factors and exhibits highly nonlinear and dynamic characteristics,which make it difficult to forecast.Existing forecasting methods have many limitations,such as insufficient extraction of load time series features and incomplete consideration of load influencing factors,which make it difficult to further improve the forecasting accuracy.To this end,this paper proposes an improved virtual power plant load forecasting method.This method is based on the multiple scale nested long short-term memory(MSNLSTM)neural network and constructs a multi-level long short-term memory network to extract the patterns of load sequences at different time scales,so as to deeply learn the inherent periodicity and correlation characteristics of the load.At the same time,external factor datas are introduced as the input of the network to enhance the modelling ability of load influencing factors.Experiments results show that compared with a single long short-term memory network and traditional forecasting methods,the improved model can improve the accuracy of day-ahead and week-ahead load forecasting.