Optimization Of A Facility Layout Based On Variational Autoencoder

In the context of carbon neutrality,China has intensified its efforts to adjust the energy structure,aiming to accelerate the deployment of clean energy and strive to realize the lowcarbon,green and sustainable national industrial chain.As one of the clean energy,nuclear power is an important part of it.Therefore,in the 14 th Five-Year Plan,the expectation of 70 million kilowatts of nuclear power in operation is clearly stipulated.The comprehensive construction of nuclear power plants has put forward higher requirements for the design of nuclear power plants to ensure the safety,reliability and efficiency of the operation of nuclear power plants.The layout design of nuclear reactor equipment is the key point of the overall design of nuclear power plant.It refers to the layout of equipment driven by multiple constraints in a closed space,with full consideration of the attributes and characteristics of all levels of equipment and various constraints among the equipment.In layout design of nuclear reactors,need to face to cloth object when the constraint conditions such as more difficult,plus according to different layout requirements often need to design a proprietary algorithm to generate the corresponding constraints of equipment layout,this makes it harder to work the design of nuclear reactor,the serious influence the design efficiency,increasing the design cost greatly.In view of the above problems,this paper aims to propose an automatic generation technology of nuclear reactor equipment layout based on artificial intelligence technology,in an attempt to enable artificial intelligence to learn and master the design experience of nuclear reactor equipment layout and quickly generate equipment layout that meets the corresponding constraints according to different design requirements.This paper elaborates from the following three aspects:(1)In this paper,the mathematical model of equipment layout scene is established,and the constraints such as symmetry,proximity,low center of gravity,boundary constraint and overlap constraint are discussed.Furthermore,the quantitative calculation of constraint conditions can be used as the quantitative evaluation index of network generation layout results or as the loss item of network training.(2)Furthermore,a device layout generation network based on structured deformable grid is constructed.The model takes a single equipment in a nuclear reactor as the coding unit,and codes the constraints related to the equipment,other equipment associated with the equipment and deformation information of the equipment.The same coding is carried out for other equipment and input into the equipment layout generation network for training.The trained network can complete the automatic generation of nuclear reactor equipment layout.(3)Aiming at the incompleteness of the equipment information coding expression,the undirectability of the layout generation process and the undirectability of the output equipment layout data of the equipment layout production network based on structured variable grid,the equipment layout generation network based on conditional variational autoencoder was proposed.The network takes the overall nuclear reactor equipment layout as the coding unit,converts the real equipment layout into a scene graph containing various constraints,and uses graph convolutional neural network to further process the input scene graph.The device layout features are learned by predicting the boundary boxes of objects in the scene graph.Based on the learned device distribution features and the scene graph as the condition,the generated device layout meeting the requirements of the given evaluation index is quickly generated.