Research on computational intelligent methodologies and optimization of wireless sensor and actuator networks

Yamin Han 수원대학교 2021 국내박사

With the development of multi-robots and wireless sensor network technologies, a new distributed wireless sensor and actuator network (WSANs) has emerged. This kind of network can not only sense the environmental information, process event data, but also interact with the monitoring environment and perform corresponding tasks to change the environment. WSANs are widely used in various fields with a high degree of convenience, flexibility, and strong robustness, such as environmental monitoring, smart cities, and medical health. However, there are some issues in WSAN systems such as limited energy, limited number of actuators, limited communication capacity, long response time, and strong dynamic, etc. These issues seriously affect the performance and service quality of WSAN systems. In addition, the large-scale and dynamic characteristics in WSAN increase the complexity and uncertainty of these issues faced by WSANs. This makes that solving these issues frequently are difficult and cost many computational time especially for manual processing. The development of computational intelligence (CI) provides the effective solutions to the issues faced by WSANs. CI draws on the ideas of bionics to solve the problems of large scale, strong complexity, multiple constraint, non-linearity, and uncertainty in the process of industrial production and engineering practice. It has the advantages of self-organization, self-adaptive, parallelism, and fault tolerance, which provides the alternative solutions for WSANs. It can not only find an optimal solution within an acceptable calculation time, but also provide a satisfactory solution. This study researches on computational intelligent methodologies and optimization of wireless sensor and actuator networks, aiming to solve the issues faced by WSANs such as limited energy, limited communication capacity, and long response time, etc. Additionally, the related technologies of CI are researched and improved to solve these issues in WSANs and improve the performance on optimizing WSANs. In detail, the main works are as follows: Maximize the coverage of sensor nodes to the monitoring region In WSAN systems, sensor nodes communicate with each other by wireless links, and exchange the collected environmental information, so as to complete the coverage of the whole monitoring region. How to use a certain number of sensor nodes to achieve the maximum coverage of the monitoring region under the premise of maintaining network connectivity is a key problem faced by WSAN configuration. This study proposes an optimization method of sensor nodes deployment based on an improved differential evolution algorithm, which effectively maximizes the coverage of the monitoring region by the sensor nodes and maintains communication between sensor nodes. Optimizing actuators deployment using hierarchical intermittent communication particle swarm optimization Network lifetime refers to the effective operation time of WSAN system, and it is the most important standard to evaluate service quality. This study optimizes the deployment issues of actuators associated with the sensor nodes coverage and the overall energy consumption of sensor nodes using a proposed hierarchical intermittent communication particle swarm optimization. This can reduce the number of hops of communication between sensor nodes and avoid energy holes, thereby extending the lifetime of the network and reducing data transmission packet loss rate. An energy balanced cluster-routing protocol using particle swarm optimization with five mutation operators In the previous studies, multiple actuators are deployed in distributed sensor nodes to reduce the number of communication hops between sensor nodes and sink, thereby prolonging the network lifetime. However, the unbalanced number of sensor nodes between clusters leads to unbalanced energy consumption between cluster heads, thereby shortening the network lifetime. Additionally, the determination of the number of clusters has an important impact on the performance of the WSAN system. This study proposes an energy balanced cluster-routing protocol using particle swarm optimization for WSANs. Five mutation operators are also specially proposed to improve the performance of PSO in optimizing the clustering of sensor nodes. The proposed protocol can dynamically determine the number of clusters and group sensor nodes into clusters evenly. So as to balance the energy consumption of sensor nodes and prolong the network lifetime. Partition-based automatic scheduling scheme for balancing energy consumption of sensor nodes and task allocation of mobile collectors Data gathering delay affects network response time, especially for time sensitive applications such as fire monitoring. Response time is particularly important. This study proposes a novel partition strategy and automatic scheduling scheme for balancing energy consumption and task allocation for WSAN. The partition strategy is proposed to uniform the data load, which can balance the energy consumption of sensor nodes and prolong the network lifetime. Then, a task allocation method based on genetic algorithm is proposed to optimize the number of collectors and uniformly assign tasks to collectors. Additionally, the trajectory of each collector is optimized by ant colony optimization algorithm. These can uniform the time consumption and find the shorter travel path of collectors, thereby reducing data gathering delay.

Particle swarm optimization via global-local scheme for structural design of aerospace system

강희진 서울대학교 대학원 2015 국내석사

To ensure the structure to be safely maintained in flight phase, it is important to design and optimize the structures including aerospace system such as flight wing, launch vehicle and satellite. The structures of aerospace system are usually configured with stiffener and panel. Since panel is surrounded by stiffener, it can be seen that local characteristics in a particular analysis such as buckling analysis. Because load path does not change significantly, panel is suitable to perform the local buckling analysis because the buckling mode is present in the local area surrounded by the stiffener. This buckling analysis is time consuming work in the structural design optimization. To perform the structural design optimization effectively, it is essential that separate the optimization problem into the global optimization problem and the local optimization problem. In this study, the global-local structural optimization problem was configured for the effective optimization of the structures including aerospace system structures. Particle swarm optimization algorithm which is useful for structural design optimization was used. To apply the global-local scheme into particle swarm optimization algorithm, optimization module was developed. This module is called the global-local PSO module. This module was constructed with three interface dialog. One is for setting optimization problem. And other one is for setup optimization environmental parameter. Third interface dialog is to start optimization and monitoring. All of these functions were realized in DIAMOND/IPSAP which is being developed by Aerospace Structures Laboratory in Seoul National University. To evaluate a performance of the particle swarm optimization algorithm using the global-local PSO module, Stiffened shell box and launch vehicle models were designed and were optimized by the global-local PSO algorithm. In case of stiffened shell box, local static buckling analysis was performed. Critical buckling load was used for constraint of the local optimization. Last example was the optimization of the launch vehicle. The significance of this study was that it was possible to faster optimization by using the global-local PSO algorithm which is appropriate approach for the structures which are able to separate into the global and local area. Also, using the computer aided engineering including the high performance solver named IPSAP and the optimization module with in-house pre/post tools, the time and effort for finding the optimal design variables which are used in the aerospace system structures including flight wing box, launch vehicle, etc are decreased efficiently.

Study on Novel Design of Methodologies for Computational Intelligence and Construction of Models for Cement Hydration

Zhang, Liangliang 수원대학교 2022 국내박사

시멘트는 국민 경제 발전에서 없어서는 안 될 역할을 한다. 그러나 시멘트 재료의 연구와 디자인은 많은 문제에 직면하여 그 발전을 제약했다. 가장 주의해야 할 것은 시멘트의 수화와 경화 과정 중의 복잡한 화학 반응과 물리적 변화로 인해 시멘트 생산 과정을 정확하게 통제하기 어렵다는 것이다. 이것 또한 시멘트 수화 연구의 난이도를 증가시켰다. 고성능 겔 재료를 설계하기 위해서는 시멘트의 수화 메커니즘을 근본적으로 연구할 필요가 있다. 선진적인 연구 도구를 사용하여 데이터를 수집하여 미시적인 척도에서 시멘트의 수화를 연구하는 것이 가능하게 한다. 특히 현미층 분석상 (μCT) 도입은 시멘트의 수화 구조 변화를 미시적 잣대에서 보여줄 수 있어 시멘트의 수화를 깊이 있게 이해하는 데 도움이 된다. 그러나 미시적 구조 이미지를 바탕으로 하는 시멘트 수화 모델링은 회귀, 분류와 최적화 등 다양한 임무와 문제에 직면하고 있다. 계산 지능의 이론과 방법은 시멘트 수화 연구에 지원을 제공할 수 있다. 그들은 주로 학습 시스템, 모호 논리, 진화 알고리즘, 단체 지능과 혼합 시스템을 포함하여 각종 유형의 임무를 처리한다. CI 를 시멘트 수화 연구에 더욱 잘 응용하고 CI모델의 성능을 향상시키기 위해 새로운 CI 이론모델도 설계했다. 전체적으로 본 연구는 새로운 CI 이론 모델을 제시하고 이를 수화 미시적 구조 시간 서열 이미지를 바탕으로 하는 시멘트 수화 모델링 연구에 응용하여 시멘트의 수화를 깊이 있게 이해하고 고성능 시멘트 디자인에 지원을 제공한다. 본 연구의 연구 내용과 공헌은 주로 두 가지 방향으로 나뉘는데 그것이 바로 CI 이론 모델과 시멘트 수화 모델이다. 이런 상황에서전체 연구내용은 다음과 같다. 계산 지능의 이론 모델: (1) 가우시안 혼합 모델(GMM-PRBFNNs) 을 사용하는 다항식 기반 방사형 기저 함수 신경망은 퍼지 규칙의 전제를 구성하기 위한 멤버십 함수를 생성하기 위해 GMM 을 사용하여 퍼지 신경망의 성능을 향상시키기 위해 제안됩니다. 이론적으로 GMM은 모든 유형의 분포에 적응할 수 있다. 또한 퍼지 규칙의 결과에 대해 4가지 유형의 다항식이 은닉층과 출력층 사이의 가중치로 간주됩니다. 이것은 데이터의 복잡한 비선형 본질을 파악하는 데 도움이 된다. 이러한 기능은 GMM-PRBFNNs 의 근사화 기능을 강화합니다. (2) 이중 구조를 가진 개선형 입자군 최적화 알고리즘(PPSO)을 제시했다. PPSO 에서 떼는 먼저 계층적 승격 구조(HPS)로 구성된 여러 개의 독립적인 하위 개체군으로 나뉘며, 이는 병렬로 최적을 검색하기 위해 각 계층의 하위 개체군을 보호합니다. 우수한 입자를 저층 차자종군에서 고층 차자종군으로 끌어올릴 수 있도록 단방향 통신 전략과 고기능 연산자가 한층 더 증진시긴다. 또한 HPS의 각 하위 집단에서 입자에 대해 계층적 다중 척도 최적이 구성되며, 여기서 각 입자는 서로 다른 척도의 최적 집합을 합성할 수 있습니다. HPS는 희망 구역으로만 날아가고 적응도가 낮은 입자가 전체 입자군과 경쟁하지 않도록 보호한다. 또한 이중 계층 구조는 검색의 다양성을 증가시킵니다. (3) 데이터 재구성(DR 기반 FPSL)에 기반한 퍼지 파티션 자체 학습이라는 새로운 퍼지 파티션 기술을 도입하는 인코더 및 디코더를 사용한 퍼지 파티션 학습을 기반으로 하는 새로운 데이터 재구성 기반 퍼지 추론 네트워크(DR-FIN)가 제안됩니다, 입력 공간을 인코더와 디코더를 반복적으로 훈련하여 데이터 재구성을 실현할 수 있는 파티션 공간으로 분할합니다. 간단한 신분 재구성을 방지하고 학습과 범위화 능력을 향상시키기 위해 DR 기반의 FPSL 과정에서 L2-범수 정규화, 인코더 희소 벌칙, 소음 제거 학습과 소량의 사다리 하락을 포함한 다양한 학습 전략을 실현했다. 그리고 DR-FIN은 최소 2승 추정과 L2-범수 정규화를 응용하여 후부 다항식의 계수를 훈련한다. (4) 퍼지 신경망 개발을 위해 확률 분포 및 구조 최적화(AFC-PDSO) 기반 적응형 퍼지 분류기가 제안됩니다. 각 클러스터에 속하는 패턴의 멤버십 정도는 가우스 혼합 모델의 도움으로 확률 분포를 결합하여 측정됩니다. 그 밖에 서로 다른 모호한 규칙에 대해 서로 다른 유형의 결과 다항식을 응용했다. AFC-PDSO의 구조는 초기 중심, 후속 다항식 유형과 L2-범수 정규화항의 파라미터를 포함하고 경쟁군 최적화 알고리즘으로 최적화했다. 시멘트 수화 모형: (1) 성능의 근사성과 이해에 있어서, GMM PRBFNNs 를 이용하여 시멘트를 수화하는 과정에서 시멘트의 압력 저항 강도와 3차원 미시적 구조 이미지 특징 간의 모호한 관계를 구축했다. 그레이스케일 직사각형도와 그레이스케일 공생 행렬의 특징치를 이용하여 시멘트의 수화 미시 구조를 묘사하는 3차원 이미지 특징으로 삼아 성분의 주파수와 배열을 통계처리하고 시멘트의 수화 정도를 잘 반영할 수 있다. 이것은 처음으로 이미지 자체에서 시멘트의 수화 과정에서 압력 저항 강도와 3차원 미시적 구조 이미지 특징 간의 모호한 관계를 연구한 것이다. 그 밖에 양호한 해석성을 가진”if-then” 모호한 규칙도 개선하여 처음으로 3차원 미시적 구조와 시멘트의 강도 간의 관계를 나타냈다. (2) 데이터 처리 및 미세 구조 진화에서, 시멘트 페이스트의 4D 미세 구조 이미지가 빠르게 구성됩니다. 등록 속도를 높이고 정확도를 향상시키기 위해 다중 요소 다중 계층 PSO를 사용한 새로운 부분 정보 등록 방법이 제안됩니다. 시멘트 페이스트의 구성된 4D 미세 구조 이미지는 재료 과학자가 현장에서 수화 과정을 연구하는 데 도움이 될 수 있습니다. (3) 구조 식별에서 자가 학습 및 계층적 클러스터링을 기반으로 시멘트 수화 μCT 이미지에 대한 적응형 3D 이미지 분할 방법이 제안됩니다. 서로 다른 단계, 완전히 혼합된 단계 및 시멘트 수화 이미지의 복잡한 경계의 낮은 대비를 극복하고 시멘트 데이터의 더 나은 표현을 얻기 위해 희소 자동 인코더를 사용하여 3D 시멘트 수화 μCT 이미지 자체에서 특징 추출기가 자동으로 학습됩니다. 그런 다음 시멘트 수화 μCT 이미지는 샘플을 더 잘 분리할 수 있는 거리 메트릭이 적응적으로 선택되는 다중 거리 메트릭을 사용하여 계층적 클러스터링으로 분할됩니다. (4) 성능 등급 평가에서 DR-FIN을 사용해 시멘트의 수화 강도 등급을 평가하는 비파괴적인 방법을 개발했다. 일반적으로 압력 저항 강도는 시험기를 사용하는 물리 실험에 의해 확정되는데, 이것은 파괴적일 뿐만 아니라, 시멘트를 분쇄해야 하기 때문에 연속적이지 않다. 시멘트의 미시적 구조는 시멘트의 수화 상태와 미시적 척도에서 압력 저항 강도와 관련된 정보를 직접 반영할 수 있음을 감안하여 본 연구는 미시적 구조 이미지의 강도 평가 등급을 바탕으로 무손상 검측 방법을 제공하였다. (5) 성능 분류에서 시멘트 수화 강도 등급은 AFC-PDSO 를 사용하여 분류됩니다. 제안된 AFC-PDSO 는 가우스 혼합 모델의 도움으로 확률 분포를 결합하여 구성원 정도를 측정하고 초기 중심, 각 퍼지 규칙의 결과 다항식 유형 및 L2-범수 정규화 항의 매개변수를 포함하는 구조를 최적화합니다. 경쟁적인 스웜 옵티마이저. 따라서 서로 다른 유형의 퍼지 규칙이 서로 다른 강도 등급에 대해 적응적으로 학습될 수 있고 좋은 성능을 보여줍니다. Cement plays an indispensable role in the development of the national economy. However, the research and design of cement materials face many problems that restrict its development. Most notably, the complex chemical reactions and physical changes during cement hydration and hardening make it difficult to precisely control the cement production process. This also increases the difficulty of cement hydration research. In order to design high-performance cementitious materials, it is necessary to study the mechanism of cement hydration in essence. Data acquisition using advanced research tools makes it possible to study cement hydration at the microscopic scale. In particular, the introduction of microtomography (μCT) can exhibit the structural evolution of cement hydration at microscale, which is helpful to understand cement hydration in depth. Nevertheless, cement hydration modeling based on microstructural images faces many different tasks and problems that cover regression, classification, and optimization, etc. The theories and methods of computational intelligence (CI) can provide support for the research of cement hydration. They mainly include learning systems, fuzzy logic, evolutionary algorithms, swarm intelligence, and hybrid systems, thereby handling a wide range of types of tasks. In order to better apply CI to various problems in cement hydration research and improve the performance of CI models, some new theoretical models of CI are also designed. Overall, this thesis presents some new theoretical models of CI and applies them to the hydration modeling research of cement based on the time series images of hydration microstructure to understand cement hydration in depth and provide support for high-performance cement design. The research contents and contributions in this study are mainly divided into two directions: theoretical models of CI and hydration modeling of cement. In this case, they are summarized as follows. (a) Theoretical models of computational intelligence: (1) Polynomial-based radial basis function neural networks with Gaussian mixture model (GMM-PRBFNNs) are proposed to improve the performance of fuzzy neural networks by using GMM to generate membership functions for constructing the premise of fuzzy rules. In theory, GMM can fit any type of distribution. Moreover, four types of polynomials are considered as the weights between the hidden layer and the output layer for the consequent of fuzzy rules. This helps to capture the complex nonlinear nature of data. These characteristics enhance the approximation ability of GMM-PRBFNNs. (2) A novel promotive particle swarm optimizer (PPSO) with double hierarchical structures is proposed. In PPSO, the swarm is first divided into multiple independent subpopulations organized in a hierarchical promotion structure (HPS), which protects subpopulation at each hierarchy to search for the optima in parallel. A unidirectional communication strategy and a promotion operator are further implemented to allow excellent particles to be promoted from low-hierarchy subpopulations to high-hierarchy subpopulations. Furthermore, in each subpopulation of HPS, a hierarchical multi-scale optimum is constructed for particles, in which each particle can synthesize a set of the optima of its different scales. The HPS can protect particles that just fly to promising regions and have low fitness from competing with the entire swarm. Also, the double hierarchical structures increases the diversity of searching. (3) A novel data reconstruction-driven fuzzy inference network (DR-FIN) based on learning fuzzy partitions with encoders and decoders is proposed, which introduces a new fuzzy partition technology, named fuzzy partitions self-learning based on data reconstruction (DR-based FPSL), to divide the input space into partition space that can realize data reconstruction by iteratively training encoders and decoders. To prevent simple identity reconstruction as well as enhance learning and generalization abilities, multiple learning strategies including L2-norm regularization, sparsity penalty on encoders, denoising learning, and mini-batch gradient descent are implemented during the DR-based FPSL. Then, DR-FIN applies the least square estimate and L2-norm regularization to train the coefficients of consequent polynomials of consequent part. (4) Adaptive fuzzy classifier based on probability distributions and structure optimization (AFC-PDSO) is proposed to develop fuzzy neural networks. The membership degrees of patterns belonging to each cluster are measured by engaging probability distributions with the aid of Gaussian mixture model. Moreover, different types of consequent polynomials are applied for different fuzzy rules. The structure of AFC-PDSO including initial centers, types of consequent polynomials, and parameter of L2-norm regularization term, is also optimized with competitive swarm optimizer. (b) Hydration modeling of cement: (1) In performance approximation and understanding, the fuzzy relationships between cement compressive strength and 3D microstructural image features are built for cement hydration using GMM-PRBFNNs. The eigenvalues of gray-level histogram and gray-level co-occurrence matrix are taken as the 3D microstructural image features to describe the microstructure of cement hydration because they can count the frequency and arrangement of compositions, which well reflects the degree of cement hydration. This is the first time that the fuzzy relationships between compressive strength and 3D microstructural image features are studied for cement hydration from the image itself. A collection of ‘‘if-then’’ fuzzy rules with good interpretability is also refined to express the relationships between the 3D microstructure and strength of cement for the first time. (2) In data processing and microstructural evolution, the 4D microstructural images of cement paste are rapidly constructed. A novel partial information registration method with multi-factor multi-layer PSO is proposed to speed up registration and improve its accuracy. The constructed 4D microstructural images of cement paste can assist material scientists in studying the hydration process in situ. (3) In structure identification, an adaptive 3D image segmentation method is proposed for cement hydration μCT images based on features self-learning and hierarchical clustering. In order to overcome the low contrast of different phases, fully mixed phases and complex boundary of cement hydration images and acquire better representations of cement data, a feature extractor is automatically learned from the 3D cement hydration μCT images themselves using sparse autoencoder. Then, cement hydration μCT images are segmented by hierarchical clustering with multiple distance metrics, in which the distance metric that can better separate the samples is adaptively selected. (4) In performance grade estimation, a non-destructive method for estimating strength grade is developed for cement hydration using DR-FIN. Generally, the compressive strength is determined by physical experiments using testing machines, which is not only destructive but also not continuous because they need to crush the cement. Considering that the cement microstructure can directly reflect the state of cement hydration and the information related to compressive strength at microscale, this study estimates the strength grade based on microstructure images, which gives a non-destructive method. (5) In performance classification, strength grades of cement hydration are classified using AFC-PDSO. The proposed AFC-PDSO measures the membership degrees by engaging probability distributions with the aid of Gaussian mixture model and optimizes its structure including the initial centers, the type of the consequent polynomial of each fuzzy rule, and the parameter of L2-norm regularization term with competitive swarm optimizer. Therefore, different types of fuzzy rules could be adaptively learned for different strength grades and shows a good performance.

Optimal Placement and Sizing of Distributed Generation in Smart Grid using Particle Swarm Optimization

SYED MUHAMMAD ARIF 성균관대학교 일반대학원 2016 국내석사

Optimal Sizing and Placement of Distributed Generation in Smart Grid using Particle Swarm Optimization There are many technical, enviromental, economical and social advantages of placing Distributed Generation (DG) in Smart Grid. Technical benefits such as, reduce the system power loss (Active/ Reactive power loss), improve the voltage profile, postponing system upgrades, enhancing system reliability and continuity of service. By bringing the Green/Clean energy also known as renewable energy such as Solar/Wind based Distributed Generation, the amount of CO2 emission in the environment can be reduce more effectively. Therefore, such kinds of power generation are more environment friendly and economical. However, the practical application of DG unit in distribution network demonstrates difficulties in various aspect such as social, economical and political factors that affect the final optimal solution. The objective of this research is to find the optimal location and size of DG unit in radial distribution systems to minimize the active power loss and voltage profile improvement. In addition it also reduce the local trapping issue of particle into local minima which is one of the biggest issue in Particle Swarm Optimization. There are many ways of calculating the size and location of DG unit in smart Grid. For example, Analytical, Hit and Trial method and heuristic approach. The Analytical and heuristic approaches have their own limitations. For example, the Analytical Approach cannot be used for large system, it can only give good results for small system size also when the system size increase, the time for obtaining the optimal location increase significantly. On the other hand, only using heuristic approaches have some problems for obtaining the optimal solution. For example, the particles may trap in to local minima and could not find the best solution. To overcome these two issues (number of iterations & Local trapping). We proposed Hybrid Particle Swarm Optimization algorithm which is the combination of both the PSO and modified analytical method (2/3rd rule). In Standard PSO, the particles are randomly place from start to the end node of the feeder and find the optimal location. In our proposed algorithm instead of searching the optimal location and size of the DG randomly. We only allow the PSO algorithm to search the DG size and location near the 2/3rd of total load on the feeder. The novelty of our proposed algorithm is search area reduction which reduce the local trapping issue of particles and also consume less iterations for obtaining the optimal location. To prove the importance of our proposed algorithm we tested it on one practicle system, KEPCO Distribution System and one IEEE standard test system, IEEE-33 bus Distribution System.

Modeling and Optimization Methodologies for Natural Gas Liquefaction Process

칸모드샤리크 영남대학교 대학원 2011 국내석사

Due to clean burning features and the ability to meet tough environmental requirements, the demand of LNG has increased considerably and the projection reflects a continued increase for the next several years [1]. Natural gas is often found at remote locations and to bring it to the world market liquefaction is required. It can reduce the volume of natural gas to 600th of the original volume and eases transportation issue. The Liquefaction of natural gas in a mixed refrigerant system is an energy demanding process. A lot of work has been wasted due to irreversibility in the process or the process in not working at optimal operation where it can save considerable amount of energy. The key physical parameters that perform the dominant role in affecting the overall performance of LNG plant are refrigerant composition, refrigerant flow rate, suction and evaporation pressure and the amount of mixed refrigerant vaporization. These variables should be adjusted to optimize the overall operation. The adjustment of one of these variables will have consequence for the other variables due to highly nonlinear nature of interaction existed between the variables. This thesis investigates various optimization methodologies for finding the optimal condition in Natural Gas liquefaction process where it can saves considerable amount of energy. Gradient-based optimization methods like Nonlinear Programming and Multi-star as well as gradient free methods like Pattern Search, Genetic Algorithm, Particle Swarm optimization were investigated in particular. The NG liquefaction plant was modeled in UniSim Design commercial process plant simulator and the model was optimized for compression energy. First, the optimization was carried out within the built-in optimizer of simulator. The inefficiency of handling the constraints within built-in optimizer led the optimization to be carried out in other platform. The optimization was then carried out in the Matlab by connecting it with the simulator using COM functionality. The optimization was carried out using both gradient as well as non-gradient methods and Particle Swarm Optimization proves to be the most successful algorithm among all in terms of energy savings. Nonlinear programming method was proved best if sought near optimal solution whereby, PSO can be used to find the optimal without any initial guess. The results were deducted from simulation results by performing case studies. 청정 연료로서 그리고 엄격한 환경 규제에 적용 가능하기 때문에 LNG의 수요는 상당히 증가했고 이러한 추세를 반영하여 앞으로 몇 년간 지속적으로 증가할 것이다. 일부 천연가스는 실제 소비 시장과 떨어진 곳에서 발견되기 때문에 운반을 위해서 액화공정이 필요하다. 액화를 함으로서 부피가 600배 줄어들고 이는 운송을 편리하게 해준다. Mixed Refrigerant System을 이용한 액화공정은 에너지가 요구되는 공정이다. 상당량의 일이 공정 내의 비가역성과 에너지를 절약할 수 있도록 최적화된 운전을 하지 않기 때문에 낭비되고 있다. LNG 공장의 전반적 수행에 영향을 주는 주요 물리적 변수는 냉매의 조성, 냉매의 양, 흡입압력, 토출압력, 그리고 증기압력과 기화되는 혼합냉매의 양이다. 이 변수들은 공정의 최적화를 위해 반드시 조정되어야 한다. 이중 한 변수의 조정은 기존 변수들 간의 높은 비선형적 상호작용에 의하여 다른 변수들의 결과 값을 얻어낼 것이다. 본 논문에서는 상당량의 에너지를 줄일 수 있도록 하는 천연가스 액화공정의 최적 조건을 찾기 위한 방법론을 조사하였다. 변화율을 근거로 하여 Pattern Search, Genetic Algorithm, Particle Swarm Optimization과 같이 변화율에 영향을 받지 않는 방법 뿐 만 아니라 비선형 프로그래밍, multi-star와 같은 최적화 방법이 부분적으로 조사되었다. 천연가스 액화공정은 상용 공정 모사프로그램인 UniSim-Design으로 모사되었고, 압축에너지에 최적화되었다. 먼저, 공정의 최적화는 공정모사 프로그램 내의 최적화 모듈로 수행되었다. 공정모사 프로그램으로 다룰 수 없는 최적화의 경우 다른 소프트웨어를 사용하여 수행하였다. 최적화는 COM기능을 사용하여 Matlab과의 연계로 수행하였다. 최적화는 변화율에 의한 방법과 변화가없는 방법에 의하여 수행되었다. Particle Swarm Optimization은 에너지 절약에 관한 가장 좋은 알고리즘으로 증명되었다. 근접한 최적화 방법을 찾는데 있어서 비선형 프로그래밍은 가장 좋은 방법으로 입증되었고, 그 중 PSO는 어떠한 초기값에 대한 추측 없이 최적화에 사용될 수 있다. 결과적으로 case-study를 통하여 공정모사로부터 결과가 도출되었다.

Overlay mark sampling optimization for wafer area using a residual weighted sparse particle swarm optimization

신재형 Greduate School, Korea University 2021 국내석사

Photolithography is the process of printing a pattern on a wafer by passing light through a mask containing a circuit pattern. The overlay error is the difference between the center of the front and rear layers, which is measured via overlay marks within the wafer. Selecting the location of the overlay marks is crucial because it significantly affects the measurement values. In addition, the wafer is divided into shot areas, and several shot areas comprise a wafer area. Therefore, it is important to perform the overlay mark sampling in the wafer area. Recently, the application of sparse particle swarm optimization (SPSO), a distance-based selection method based on particle swarm optimization, has been studied for sampling optimization. The SPSO uses only the distance between marks and demonstrates high performance in the shot area. However, there is a limit to the improvement that can be achieved in the local overlay errors in the wafer area. In this paper, we propose a residual weighted SPSO to optimize the overlay mark sampling based on the distance and residual between marks in the wafer area. Several experiments were performed with the use of real overlay data on two different layers, and it was shown that the proposed method is superior to other algorithms.

FedPSO : Particle Swarm Optimization을 사용하여 네트워크 통신량을 감소시킨 Federated Learning

박성환 중앙대학교 대학원 2021 국내석사

연합학습(FL)은 데이터를 서버로 수집하지 않고, 데이터를 갖고 있는 분산 클 라이언트에서 직접 학습을 진행하며, 학습된 모델만을 서버로 수집하여 데이터 프라이버시를 보장 받을 수 있는 ML 학습 방법이다. 연합학습의 클라이언트는 일반적으로 통신 대역폭의 큰 제약이 존재하기 때문에 성능 향상을 위해서는 서버-클라이언트 간의 통신 성능을 증가시켜야 한다. 이 논문에서는 기존 많은 연합학습에서 사용하던 학습된 모델의 가중치를 수집하여 글로벌 모델을 업데 이트하는 FedAvg 대신 Particle Swarm Optimization 알고리즘의 특징을 활 용하여 클라이언트가 서버로 전송하는 데이터의 형태를 변경함으로써 향상된 네트워크 통신 성능을 보유한 글로벌 모델 업데이트 방법 FedPSO을 제안한다. 이 논문의 결과는 제안된 모델을 적용함으로써 네트워크 통신에 사용되는 데이 터의 양을 크게 줄였으며, 그럼에도 글로벌 모델의 정확도가 향상되었음을 보 여준다. 또한 불안정한 네트워크 환경에서도 기존 FedAvg보다 견고함을 보여 준다. Federated learning is a machine learning model that can collect only learned models on a server and ensure data privacy. Federated learning does not collect data as a server, but it proceeds directly from distributed clients with data. Since the client of Federated Learning generally has a big constraint on communication bandwidth, communication performance between server and client should be increased to improve performance. In this study, we use the characteristics of the particle swarm optimization algorithm instead of FedAvg, which updates the global model by collecting weights of learned models that were mainly used in federated learning. As a result, we propose a FedPSO model, which is a global model update method with improved network communication performance by changing the form of data that clients transmit to servers. The results of this study show that the amount of data used in network communication is greatly reduced by applying FedPSO, and the accuracy of global models is improved. It also shows that it is stronger than FedAvg in the unstable network environment.

Smart charging method of electric vehicles using selective binary particle swarm optimization

이영준 Graduate School, Yonsei University 2017 국내석사

최근 들어 화석 연료의 사용량이 꾸준하게 증가하고, 이에 따라 발생하는 다양한 환경 문제들에 대한 걱정이 늘어나고 있다. 이 문제들을 해결하기 위한 방안으로 전력 분야에서는 전기자동차 및 관련 산업이 큰 관심을 받고 있다. 하지만 전기자동차 도입의 증가는 기존의 약한 전력 계통에서의 안정성 문제 등 새로운 문제를 야기한다. 본 논문에서는 이 문제를 해결하고자 휴리스틱 알고리즘 (heuristic algorithm) 중 입자군집 최적화 (particle swarm optimization)를 활용한 새로운 전기자동차 스마트 충전 기법을 제안한다. 본 논문에서 전기자동차는 완속 및 급속 충전기에 연결되는 단순 유효 전력 부하로 활용된다. 전기자동차 배터리 용량의 특성 상 급속 충전기에 연결되는 전기자동차는 충전을 하는 데에 한 시간 미만의 시간이 소요된다. 하지만 완속 충전기에 연결되는 전기자동차는 수 시간의 충전 시간이 필요하므로, 시뮬레이션에 활용될 “충전 시간”을 설정하기 위해 집 출발 및 도착 시간, 배터리 SOC (state-of-charge) 레벨 등의 다양한 요소들을 고려한다. 또한 현실성을 반영한 확률론적 모델링을 통하여, 일일 가정 부하 프로파일과 전기자동차 부하 프로파일을 구성하고 몬테카를로 시뮬레이션 (Monte Carlo simulation)을 통해 결과를 확인한다. 본 논문에서는 제안하는 스마트 충전 기법을 활용하여 IEEE 13 node test feeder 와 IEEE 34 node test feeder 의 두 시스템에서 배전 계통 손실 최소화를 위한 최적화를 수행한다. 시뮬레이션은 OpenDSS와 MATLAB을 통해 수행하며, 전체 계통 손실과 전압위반요소 (voltage violation factor)를 최소화하는 데에 목적을 둔다. 더 나아가 전기자동차 부하가 제일 큰 노드에서, 부하가 가장 집중된 시간 대의 부하율 (load factor)과 전압 변동성 (voltage variability)을 분석하여 전체 시스템 뿐만 아니라 하나의 노드에서도 제안된 스마트 충전 기법이 효용적으로 적용됨을 검증한다. As concerns about environmental issues caused by the use of fossil fuels grow, electric vehicles (EVs) and their associated infrastructures have attracted attention from the perspective of electric systems. However, with the penetration of EVs, stability issues are also arising in existing weak systems. To solve these issues, this thesis suggests a novel smart charging method based on heuristic algorithms, especially particle swarm optimization. EVs are considered as a simple active power load connected to either a slow or fast EV charger. Whereas an EV using a fast EV charger requires less than an hour to charge its EV battery, an EV using a slow EV charger consists of various references such as home departure and arrival times and battery state-of-charge (SOC) level to build a “charging time” matrix that can be used with a simulation. To reflect realistic conditions, stochastic models comprising daily load profiles and EV load profiles are constructed. Monte Carlo simulations are performed using the stochastic models to extract simulation results. The proposed smart charging method is used to optimize distribution system loss in the case of two simulations: one using an IEEE 13 Node test feeder and the other using an IEEE 34 Node test feeder. The simulations are conducted through OpenDSS and are aimed at verifying the effectiveness of the proposed smart charging method in minimizing system losses and voltage violations. Furthermore, a load factor and voltage variability are analyzed to identify the effects of the proposed smart charging method on the node having the highest EV load.

적층제조공정을 통해 제작되는 Meta Grain Structure의 기계적 물성 제어에 관한 연구

김기범 인하대학교 대학원 2021 국내박사

Metaheuristic을 활용한 황화물계 고체 전해질 및 형광체 소재 연구

박준서 세종대학교 대학원 2023 국내석사