Showing papers by "Zhile Yang published in 2018"

Compact real-valued teaching-learning based optimization with the applications to neural network training

Abstract: The majority of embedded systems are designed for specific applications, often associated with limited hardware resources in order to meet various and sometime conflicting requirements such as cost, speed, size and performance. Advanced intelligent heuristic optimization algorithms have been widely used in solving engineering problems. However, they might not be applicable to embedded systems, which often have extremely limited memory size. In this paper, a new compact teaching-learning based optimization method for solving global continuous problems is proposed, particularly aiming for neural network training in portable artificial intelligent (AI) devices. Comprehensive numerical experiments on benchmark problems and the training of two popular neural network systems verify that the new compact algorithm is capable of maintaining the high performance while the memory requirement is significantly reduced. It offers a promising tool for continuous optimization problems including the training of neural networks for intelligent embedded systems with limited memory resources.

Iterative Learning Identification and Compensation of Space-Periodic Disturbance in PMLSM Systems With Time Delay

Abstract: This paper aims to solve a closed-loop identification problem for the space-periodic force ripple in permanent-magnet linear synchronous motor (PMLSM) systems. Conventional identification schemes use the overall error signal to update estimates. However, the error caused by mechanical vibration and measurement noise could affect and even deteriorate the identification performance. In this paper, a novel iterative learning identification method that utilizes the partial but most pertinent information in the error signal is proposed to identify the force ripple. First, the effective error signal caused by the reference trajectory and the force ripple are extracted by projecting the overall error signal to a subspace. The subspace is spanned by some basis functions selected on the basis of the physical model of the PMLSM and the sinusoidal model of the force ripple. The time delay of the PMLSM system is compensated in these basis functions. Then, a norm-optimal approach is proposed to design the learning gain. The monotonic convergence of the iterative learning identification is further analyzed. Numerical simulation and experiments are provided to validate the proposed method and confirm its feasibility and effectiveness in force ripple identification, as well as its compensation.

Complex Power System Status Monitoring and Evaluation Using Big Data Platform and Machine Learning Algorithms: A Review and a Case Study

Abstract: Efficient and valuable strategies provided by large amount of available data are urgently needed for a sustainable electricity system that includes smart grid technologies and very complex power system situations Big Data technologies including Big Data management and utilization based on increasingly collected data from every component of the power grid are crucial for the successful deployment and monitoring of the system This paper reviews the key technologies of Big Data management and intelligent machine learning methods for complex power systems Based on a comprehensive study of power system and Big Data, several challenges are summarized to unlock the potential of Big Data technology in the application of smart grid This paper proposed a modified and optimized structure of the Big Data processing platform according to the power data sources and different structures Numerous open-sourced Big Data analytical tools and software are integrated as modules of the analytic engine, and self-developed advanced algorithms are also designed The proposed framework comprises a data interface, a Big Data management, analytic engine as well as the applications, and display module To fully investigate the proposed structure, three major applications are introduced: development of power grid topology and parallel computing using CIM files, high-efficiency load-shedding calculation, and power system transmission line tripping analysis using 3D visualization The real-system cases demonstrate the effectiveness and great potential of the Big Data platform; therefore, data resources can achieve their full potential value for strategies and decision-making for smart grid The proposed platform can provide a technical solution to the multidisciplinary cooperation of Big Data technology and smart grid monitoring

An Investigation of Wind Direction and Speed in a Featured Wind Farm Using Joint Probability Distribution Methods

Abstract: Wind direction and speed are both crucial factors for wind farm layout; however, the relationship between the two factors has not been well addressed. To optimize wind farm layout, this study aims to statistically explore wind speed characteristics under different wind directions and wind direction characteristics. For this purpose, the angular–linear model for approximating wind direction and speed characteristics were adopted and constructed with specified marginal distributions. Specifically, Weibull–Weibull distribution, lognormal–lognormal distribution and Weibull–lognormal distribution were applied to represent the marginal distribution of wind speed. Moreover, the finite mixture of von Mises function (FVMF) model was used to investigate the marginal distribution of wind direction. The parameters of those models were estimated by the expectation–maximum method. The optimal model was obtained by comparing the coefficient of determination value (R2) and Akaike’s information criteria (AIC). In the numerical study, wind data measured at a featured wind farm in north China was adopted. Results showed that the proposed joint distribution function could accurately represent the actual wind data at different heights, with the coefficient of determination value (R2) of 0.99.

Mean-square error constrained approach to robust stochastic iterative learning control

Abstract: A Kalman filtering-based robust iterative learning control algorithm is proposed in this study for linear stochastic systems with uncertain dynamics and unknown noise statistics. Firstly, a learning gain matrix is designed for the nominal case by minimising the trace of the mean-square matrix of the input tracking error. Theoretical results show that the proposed algorithm guarantees not only the asymptotic but also monotonic convergence of the input tracking error in the mean-square error sense, especially when random noises are Gaussian distributed the proposed algorithm is further proved to be asymptotically efficient. In addition, a new mean-square error constrained approach is presented in designing the robust learning gain matrix, taking into account model uncertainties. A sufficient condition is provided such that the mean-square matrix of the input tracking error is constrained within a predesigned upper bound which can monotonically converge to zero. Finally, numerical examples considering both structured and unstructured model uncertainties are included to illustrate the effectiveness of the proposed algorithms.

A Novel Binary Jaya Optimization for Economic/Emission Unit Commitment

Abstract: Economic unit commitment is a mix-integer large scale optimization problem calling for powerful and efficient tools. On the other hand, environmental impact related to the power generation is attracting increasing attentions due to the global warming trend and urgent calls for sustainable energy development. In this paper, the dual objectives of economic and emission unit commitment is converted into a single objective problem. For solving this, a novel binary Jaya optimization is proposed and integrated with lambda iteration method. The proposed binary Jaya method is inspired by the Jaya evolution and generates binary bits from a v-shape transfer function. Numerical study demonstrates the significant improvement of the binary Jaya in regarding the convergence speed for solving unit commitment problem. The solution distributions of the both objectives also show the effective of the proposed methods.

Robust control design for the trajectory tracking of a quadrotor

Abstract: The paper proposes two different controllers for the inner and outer loop of the quadrotor. The dynamic mathematical model of a quadrotor is derived based on Euler-Lagrange formulation. The outer loop controller design is verified with two controllers. One is the D (derivative) controller with a derivative filter and the second one is a PD (proportional-derivative) controller with derivative filter proposed for the position tracking of the quadrotor, where the advantage of the proposed controllers include easiness of implementation, higher stability and less dependence on the system model. The integral sliding-mode controller is proposed to obtain the desired attitude angles for the quadrotor tracking and its advantage is unaffected by the system model errors. The designed controllers can stabilize the Euler angles and make the quadrotor to move $(x,y,z)$ positions to their desired values. The validation of the results has been tested in MATLAB simulation and exhibits adequate performance.

Residual Blocks PointNet: A novel faster PointNet framework for segmentation and estimated pose

Abstract: Given recent advances in Segmentation of Convolutional Neural Networks (CNNs), this paper aims to propose a more efficient structure which directly consumes point clouds for segmentation and estimated pose. More specifically, a novel Residual Blocks PointNet is proposed providing a fast framework taking point sets as input and predicting 3D object part segmentation and 3D pose. The network of the proposed structure has been established composed of two subnet works: a key branch for 3D object part segmentation and the other branch for spatial transform to predict a 3D affine matrix. The major branch contains more residual blocks, which encapsulate shortcut connects with specified layer numbers, growth rate and conv (1*1)-bn-relu structure. The key point is the decrease of each level of network computing and the reuse of feature maps. The other is a parallel classification network for estimated pose with share portion weight except 3 groups of full connected layers. Empirically, Residual Blocks PointNet shows faster rate of convergence and acceptable performance.

The College Compus Energy Monitoring Platform for Artificial Intelligence Application

Abstract: In this paper, an energy consumption monitoring platform for the college compus has been developed. The implemented platform can realize the distributed monitoring and centralized control & management for the real-time energy and water consumption data for four-level measurement. It can provide data analysis, data storage, information remote transmission and intuitive display to the administrative control center and personnel. Furthermore, it can realize the water metering balance and appropriate public air-conditioner energy adjusting consumption. The water and electricity consumption of the student dormitories and office buildings can be monitored, remotely measured and intelligently paid, possessing the antitheft function. Besides, the specific energy APPs are also developed to control the terminals remotely via the mobile phone. The designed energy monitoring system has been implemented in the compus successfully.

Human localization and tracking system based on multiple depth cameras in VR Scene

Abstract: Reliable and accurate human localization and tracking systems is one of key technologies in virtual reality (VR). However, the traditional 2D object detection method is vulnerable to the changes in illumination, complex background, object occlusion, shadow interference and other factors, leading to poor robustness in the real world detections. To solve this problem, this paper proposes a human positioning and tracking system based on advanced multi-depth cameras. Multiple 3D TOF intelligent depth cameras are utilized to work coordinately, combining the Histograms of Oriented Depth (HOD), to extract human head and shoulder features. Then, a SVM classifier is used to achieve the target detection, followed by the improved Cam-shift algorithm for tracking real-time single target. Experimental results show that the system has high real-time performance and accuracy, providing a competitive solution for VR applications.

Compact Neural Modeling of Single Flow Zinc-Nickel Batteries Based on Jaya Optimization

Abstract: As a novel family member of the redox flow batteries (RFBs), the single flow zincnickel battery (ZNB) without ion exchange membranes has attracted a lot of interests in recent years due to the high charging and discharging efficiencies. To understand the electrical behaviour is a key for proper battery management system. Unlike the electrochemical mechanism models and equivalent circuit models, the neural network based black-box model does not need knowledge about the electrochemical reactions and is a promising and adaptive approach for the ZNB battery modelling. In this paper, a compact radial basis function neural network is developed using a two-stage layer selection strategy to determine the network structure. While Jaya optimization is utilized to determine the non-linear parameters in the selected hidden nodes of the resultant RBF neural network (RBF-NN) model. The proposed method is implemented to model the ZNB to capture the non-linear electric behaviours through the readily measurable input signals. Experimental results manifest the accurate prediction capability of the resultant neural model and confirm the effectiveness of the proposed approach.

A Novel 3D Head Multi-feature Constraint Method for Human Localization Based on Multiple Depth Cameras

Abstract: Up to date, the majority of existing spatial localization methods is based on visual positioning methods and non-visual positioning methods. In the vision-based positioning method, the traditional 2D human detection method are vulnerable to tackle with environmental changes including illumination, complex background, object occlusion, shadow interference and other factors, due to which the algorithm is less robust and difficult to achieve accurate target positioning. In respects to 3D human positioning, binocular vision or multi-vision approaches have been widely used to acquire depth information. The complexity of the algorithm is high, and the detection range is limited. To deal with this, a spatial location method based on multiple depth cameras is proposed in this paper. Multiple 3D-TOF depth cameras are jointly used to directly obtain depth information. Histograms of Oriented Depth (HOD) features are then extracted and trained to find the human head and shoulder region. Moreover, Spatial Density of Head (SDH) and the Convexity and Square Similarity of Head (CSSH) features are combined to determine the human target. Finally, the positioning data of multiple cameras are determined by using Nearest Center Point (NCP) to obtain the final human body positioning information. Experimental results show that the proposed method can not only obtain higher recognition rate and positioning accuracy, but also enable a larger detection range, meeting the needs of large-scale spatial positioning.

Detection of Transient Signals in Smart Grid Using Artificial Neural Network Modeling and JAYA Optimization

Abstract: In order to solve the power quality issues facing by Smart Grid with the rapidly growing of renewable energy sources and types, a data-driven approach to detect transient evens using Artificial Neural Network modeling and optimization is proposed in this paper. This approach is realized by building a detection model based on the historic dataset, using statistical processing strategy and modeling methods. To be specific, a fault reconstruction stage is first designed to pre-process the group of dataset to check the synchronization of different signals and roughly detect the transients or faults that most likely happened. Then, the unsynchronized signal is further processed using a detection model. In order to improve the model accuracy without adding more inputs or neural network nodes, a recently proposed optimization method named JAYA is applied to tune the parameters of the network, with a simple one phase of evolutionary process. At last, the linear residuals are generated and modeled using partial Principal Component Analysis, and only $T^{2}$ statistic is calculated and monitored instead of the traditional two statistics, thus the new detection scheme is easier and faster to apply. The proposed method is demonstrated to be effective using real-system experiment, and it is a reliable monitoring scheme to improve the level of safety for Smart Grid.

Demand Based Bidding Strategies Under Interval Demand for Integrated Demand and Supply Management

Abstract: The penetration of renewable resources in the wholesale electricity market and the demand response in the retail market cause the demand and the supply to become more unpredictable. The ISO is hard to efficiently schedule the production and dispatch the demand. Furthermore, strategic bidding in a more competitive environment is an important problem for the generator. Forecasting the hourly market clearing price (MCP) in the day-ahead electricity market is one of essential task for any bidding decision making. But only a single predicted value of MCP cannot offer enough help for the generator to select the optimal bidding strategies. Aiming at challenge these tasks, we design a new wholesale mechanism in which the ISO declares an interval demand to the wholesale market. The interval demand is more robust than a single demand figure and enables the ISO to handle unpredictable demand under the DR programs. We also developed a forecasting model to forecast a MCP function under the interval demand and introduce the notion of confidence interval to the forecasting model. The confidence interval predicts the exact range of hourly MCP. Based on these work, the optimal bidding strategies for the generator under an interval demand is also illustrated.

A Fast Hybrid Meta-Heuristic Algorithm for Economic/Environment Unit Commitment with Renewables and Plug-In Electric Vehicles

Abstract: To tackle with the urgent scenario of significant green house gas and air pollution emissions, it is pressing for modern power system operators to consider environmental issues in conventional economic based power system scheduling. Likewise, renewable generations and plug-in electric vehicles are both leading contributors in reducing the emission cost, however their integrations into the power grid remain to be a remarkable challenging issue. In this paper, a dual-objective economic/emission unit commitment problem is modelled considering the renewable generations and plug-in electric vehicles. A novel fast hybrid meta-heuristic algorithm is proposed combing a binary teaching-learning based optimization and the self-adaptive differential evolution for solving the proposed mix-integer problem. Numerical studies illustrate the competitive performance of the proposed method, and the economic and environmental cost have both been remarkably reduced.

A Natural Scene Edge Detection Algorithm Based on Image Fusion

Abstract: Convolutional neural network (CNN) has been widely used in the edge detection areas and shown competitive results. However, with the increase of receptive fields, the convolution features in CNN gradually become rough and difficult to figure out. To tackle with the problem, a novel network is proposed in this paper, making full use of the multi-scale and multi-level information of the object to perform image-to-image prediction, and combining all distinctive convolution features in a holistic manner. Further, the effect of simply connecting the feature map is enhanced by an image fusion algorithm to improve the utilization of features. The feature maps obtained by convolutions of each layer are fused through the fusion network to obtain a more detailed feature. The improved algorithm is validated in the BSDS500 dataset and the ODS F-measure has reached 0.818, which significantly exceeds the current state-of-the-art results.