Studyon Deep Reinforcement Learning Method For Optimization Design Of Wind Turbine Parameters

Due to complexity and expense of manufacturing,complex product’s important parameters need to be optimized.The aim is to reduce the cost as far as possible on the premise that the output performance is satisfied.When facing these complex and multi-parameters optimization problems,optimal method cannot be valid for the calculation difficulty of derivative information.So some search-based heuristic algorithms(genetic algorithms(GA),particle swarm algorithms(PSO),etc.)will be more effective on this issue.These heuristic algorithms often take a lot of time to search because the search experience cannot be accumulated.The deep reinforcement learning(DRL)algorithm can accumulate the experience of agent searching the environment through training valid value function network and conduct rapid optimization in similar problems.This thesis provides a method that uses DRL to solve complex product design parameters optimization problem and reduces optimization time.Firstly,this paper establishes the theory and model of large permanent magnet direct drive wind generator’s structural parameters optimization problem.The design parameters of the wind turbine are described in detail and the calculation module based on equivalent magnetic circuit method is established.Through the analysis of wind turbine parameters optimization problem,19 important structure parameters are defined as the optimization goal.Optimization object is the total cost of effective material and the constraint conditions are five output performance index.The differently of processing different types of data and modeling calculation are analyzed as well.Secondly,this thesis innovatively proposes the method of applying DRL algorithm to parameter optimization problem.Combining the characteristics of parameter optimization problems,the state,action space and reward function are defined.The value network of neural network structure is constructed and the structure and activation function of neural network are determined through experimental analysis.Then the training and optimization process of parameter optimization problem is designed according to DRL algorithm.Thirdly,in view of permanent magnet direct-drive wind turbine structure parameters optimization problem,this thesis has realized GA,PSO algorithm and proposed DRL algorithm.Results show that the optimization result of DRL algorithm is superior to GA and PSO.DRL algorithm’s training time is much more than the former two algorithms,but well-trained value network used in the optimization process can greatly reduce the optimization time.However,DRL algorithm also has disadvantages of long training time and strong dependence on parametric structure,which needs to be used in the appropriate situation.In the end,this thesis develops a structural parameter optimization system for permanent magnet direct drive wind generator.Through demand analysis and function clarification,the parameter optimization system’s framework is designed.GA,PSO algorithm and DRL algorithm optimization interface are developed using Matlab GUI.The background calculation program is written by Matlab and Python.This system can let related technicians easily using these algorithms to optimize wind generator products.