Short-Term Probability Density Function Forecasting Of Industrial Loads Using ConvLSTM And MDN

Nowadays,the power industry is striving to transform from high-speed development to high-quality development.It is in the critical period of changing the development mode and optimizing the supply and demand structure.Power demand side management,which aims to guide customers to use electricity intelligently,improve power consumption efficiency and optimize resource allocation,has drawn great attention.Among demand-side consumers,the power energy consumption of industrial consumers accounts significant percentage and shows an increasing trend.However,industrial consumers lack a rational understanding of their complicated and redundant electricity consumption information,resulting in increased electricity costs and frequent surplus of electricity.This is not conducive to the load management of the grid.Therefore,power load forecasting is particularly important,which can help industrial consumers grasp the dynamic trend of load and power consumption behavior.But the majority of the load forecasting literature focuses on deterministic load forecasting,which does not consider the uncertainty information of industrial load.This paper proposes a probabilistic density load forecasting model comprising convolutional long short-term memory(ConvLSTM)and mixture density network(MDN).The model can directly predict probability density functions(PDFs)of future load.First,this paper uses the Pearson correlation coefficient to analyze industrial load temporal relevancy.The analysis results show that the load temporal relevancy tends to weaken with the increase of the time interval.Meanwhile,the qualitative analysis of temperature and external factors(temperature factors,calendar factors)is carried out to obtain the reasons and laws of their influence on the load,thereby improving the accuracy of load forecasting.Secondly,on the basis of retaining the correlation between loads and two influencing factors,a sliding window strategy is adopted to convert one-dimensional(1-D)data to two-dimensional(2-D)matrices to reconstruct input features,so that they can be easily and quickly used in the ConvLSTMMDN model.During the training process,it helps the ConvLSTM network to better capture the key features and hidden information in the load data.Then,this paper builds a novel hybrid model based on ConvLSTM and MDN,which aims to obtain complete statistical information about future industrial load consumption in the form of PDF.The proposed method can predict industrial loads with strong nonlinear relationship,high variability and severe uncertainty.Finally,a comprehensive case study of nonferrous metal smelting industry dataset and medical industry dataset from China and a public dataset from Ireland is conducted and compared with eight state-of-the-art models.The experimental results show that the proposed model is superior to the comparison model in both deterministic evaluation indicators and probabilistic evaluation indicators.The model in this paper has stronger robustness,better generalization performance and higher prediction accuracy.