Research And Implementation Of Power Load Analysis Method Based On Deep Neural Network

With the continuous improvement of power system informatization,the accumulated data volume of power grid develops from TB level to PB level.The study of power load big data analysis algorithm and the establishment of effective knowledge discovery model can generate huge social value and eco-nomic benefits,so it is of great significance to the development of smart power grid.For example,Improving the accuracy of power load forecasting results is related to the safety and reliable power supply of power grid,and directly affects the operation decision-making and economic benefits of power grid enterprises.While Anomaly detection can not only improve the robustness of power load prediction,but also help to reduce NTL(i.e.Non-Technical Losses).This paper is based on key technologies of big data application in intelligent power distri-bution(No.2015AA050203),which is a task of the National Hightech R&D Program.The main contents of this paper are as follows:1.We propose a power load forecasting algorithm based on deep representa-tion learning.As we can see,complex periodic,sensitivity to external fac-tors(e.g.temperature),and hidden relationship between different power types(e.g.active power and reactive power)can always present in load sequences of power systems,which bring great challenges to the forecast-ing model.Therefore,we propose NeuCast,a novel model based on deep representation learning to capture the non-linear relationship between pe-riodicity,external factors,different power types and electricity demand.Extensive experiments on 134 real-word datasets show the improvements of NeuCast over the stateof-the-art methods.2.We propose an anomaly detection algorithm under the framework of deep representation learning.Power load sequence may contain abnormal pat-terns(e.g.high temperature load,equipment failure,debugging,mainte-nance,etc.),which will directly affect the accuracy of the forecasting al-gorithm.Therefore,we propose NeuCast-AD,which learns hidden repre-sentation by deep neural networks and utilize Autoplait algorithm to detect anomaly interval in the hidden representation.In this way,the accuracy of power load sequence prediction can be further improved,the sensitivity to abnormal interval can be reduced,and the robustness of the algorithm can be improved.3.We design and impelement a visual analysis system for power load fore-casting.The task of power load analysis needs not only the support of algorithm theory,but also the implementation of big data visual analysis platform.A system based on standard flow of visual analysis is designed to help power system experts discovering unknown trends and phenomena from abundant power consumption behavior data by combining their own domain knowledge,providing reliable and understandable results.We use the visual analysis system in this paper to analyze and mine the power load data of a southern city in China,obtaining valuable conclusions,which proves the practicability of the system to some extent.