Low rank optimization for efficient deep learning:making a balance between compact architecture and fast training

Deep neural networks(DNNs)have achieved great success in many data processing applications.However,high computational complexity and storage cost make deep learning difficult to be used on resource-constrained devices,and it is not environmental-friendly with much power cost.In this paper,we focus on low-rank optimization for efficient deep learning techniques.In the space domain,DNNs are compressed by low rank approximation of the network parameters,which directly reduces the storage requirement with a smaller number of network parameters.In the time domain,the network parameters can be trained in a few subspaces,which enables efficient training for fast convergence.The model compression in the spatial domain is summarized into three categories as pre-train,pre-set,and compression-aware methods,respectively.With a series of integrable techniques discussed,such as sparse pruning,quantization,and entropy coding,we can ensemble them in an integration framework with lower computational complexity and storage.In addition to summary of recent technical advances,we have two findings for motivating future works.One is that the effective rank,derived from the Shannon entropy of the normalized singular values,outperforms other conventional sparse measures such as the?_1 norm for network compression.The other is a spatial and temporal balance for tensorized neural networks.For accelerating the training of tensorized neural networks,it is crucial to leverage redundancy for both model compression and subspace training.

Optimization on the Impeller of a Low-specific-speed Centrifugal Pump for Hydraulic Performance Improvement

In order to widen the high-efficiency operating range of a low-specific-speed centrifugal pump, an optimization process for considering efficiencies under 1.0Qd and 1.4Qd is proposed. Three parameters, namely, the blade outlet width b2, blade outlet angle β2, and blade wrap angle φ, are selected as design variables. Impellers are generated using the optimal Latin hypercube sampling method. The pump efficiencies are calculated using the software CFX 14.5 at two operating points selected as objectives. Surrogate models are also constructed to analyze the relationship between the objectives and the design variables. Finally, the particle swarm optimization algorithm is applied to calculate the surrogate model to determine the best combination of the impeller parameters. The results show that the performance curve predicted by numerical simulation has a good agreement with the experimental results. Compared with the efficiencies of the original impeller, the hydraulic efficiencies of the optimized impeller are increased by 4.18% and 0.62% under 1.0Qd and 1.4Qd, respectively. The comparison of inner flow between the original pump and optimized one illustrates the improvement of performance. The optimization process can provide a useful reference on performance improvement of other pumps, even on reduction of pressure fluctuations.

Optimization and Regulation Policy for Land Use Changes Based on Low-carbon Emission in Developed Regions of China

Land use and cover change(LUCC) is one of the important causes of the Earth’s carbon cycle imbalances resulting from failure in optimizing land use. The solution to this problem has been the hotspot of research in land and environmental science. We took 'low carbon', 'energy saving' and 'high-efficiency' as the goals of land use optimization,and integrated Markov-CA(Cellular Automaton),the Grid-Fractal model and GIS,in order to study carbon emission objective function,to establish a simulation method for land use spatial allocation optimization,to evaluate the effect of the method on carbon emissions. Regulation policy on three types of land use spatial allocation was proposed,including 'low-carbon type', 'low-carbon-economic type' and 'economic type'. We applied the method to analyze the land use spatial allocation in Taixing City of the 'Yangtze River Delta' regions in China,and obtained the following results:(i) The three optimization types would improve carbon emissions by 3. 21%,1. 80% and 0. 36% respectively in 2020,compared with 2010;(ii) The actual planning for 2020 was close to the 'low-carbon-economic type';(iii) The optimization method and regulation policy,combining local optimization and global control,could meet the sustainable multi-objective requirements for low-carbon constraints of land use spatial allocation. The result of this research could also serve as a reference for exploration into patterns of regional low-carbon land use and measures for energy saving and emission reduction.

Microstrip Low-Pass Elliptic Filter Design Based on Implicit Space Mapping Optimization

It is a time-consuming and often iterative procedure to determine design parameters based on fine, accurate but expensive, models. To decrease the number of fine model evaluations, space mapping techniques may be employed. In this approach, it is assumed both fine model and coarse, fast but inaccurate, one are available. First, the coarse model is optimized to obtain design parameters satisfying design objectives. Next, auxiliary parameters are calibrated to match coarse and fine models’ responses. Then, the improved coarse model is re-optimized to obtain new design parameters. The design procedure is stopped when a satisfactory solution is reached. In this paper, an implicit space mapping method is used to design a microstrip low-pass elliptic filter. Simulation results show that only two fine model evaluations are sufficient to get satisfactory results.

Design of multi-layered porous fibrous metals for optimal sound absorption in the low frequency range

We present a design method for calculating and optimizing sound absorption coefficient of multi-layered porous fibrous metals （PFM） in the low frequency range. PFM is simplified as an equivalent idealized sheet with all metallic fibers aligned in one direction and distributed in periodic hexagonal patterns. We use a phenomenological model in the literature to investigate the effects of pore geometrical parameters （fiber diameter and gap） on sound absorption performance. The sound absorption coefficient of multi- layered PFMs is calculated using impedance translation theorem, To demonstrate the validity of the present model, we compare the predicted results with the experimental data. With the average sound absorption （low frequency range） as the objective function and the fiber gaps as the design variables, an optimization method for multi-layered fibrous metals is proposed. A new fibrous layout with given porosity of multi-layered fibrous metals is suggested to achieve optimal low frequency sound absorption. The sound absorption coefficient of the optimal multi-layered fibrous metal is higher than the single- layered fibrous metal, and a significant effect of the fibrous material on sound absorption is found due to the surface Dorosity of the multi-layered fibrous.

Porosity Calculation of Tight Sand Gas Reservoirs with GA-CM Hybrid Optimization Log Interpretation Method

Tight sand gas reservoirs are our country’s fairly rich unconventional natural gas resources, and their exploration and development is of prime importance. Sulige Gas Field which located in the northern Ordos Basin is tight sand gas reservoirs. It is typically featured by low porosity and low permeability, and the error of porosity calculation by traditional methods is larger. Multicomponent explanation model is built by analyzing the thin slice data, and the objective function is got according to the concept of optimization log interpretation method. This paper puts the Genetic Algorithm and the Complex Algorithm together to form the GA-CM Hybrid Algorithm for searching the optimal solution of the objective function, getting the porosity of tight sandstone gas reservoirs. The deviation got by this method is lesser compared with the core porosity, with a high reliability.

Parametric optimization and performance comparison of organic Rankine cycle with simulated annealing algorithm

Taking the ratio of heat transfer area to net power and heat recovery efficiency into account,a multi-objective mathematical model was developed for organic Rankine cycle(ORC).Working fluids considered were R123,R134a,R141b,R227ea and R245fa.Under the given conditions,the parameters including evaporating and condensing pressures,working fluid and cooling water velocities were optimized by simulated annealing algorithm.The results show that the optimal evaporating pressure increases with the heat source temperature increasing.Compared with other working fluids,R123 is the best choice for the temperature range of 100-180 °C and R141b shows better performance when the temperature is higher than 180 °C.Economic characteristic of system decreases rapidly with the decrease of heat source temperature.ORC system is uneconomical for the heat source temperature lower than 100 °C.

数字经济时代管理咨询行业情报工作胜任力模型优化研究

[目的/意义]随着新一轮科技革命和产业变革不断深入,管理咨询行业日益受到重视,尤其在当前数字经济背景下,根据管理咨询行业目前的发展情况,急需对情报工作胜任力模型优化升级,为我国对管理咨询行业情报工作的有效管理提供科学参考。[方法/过程]梳理目前管理咨询企业中情报工作者胜任力不明确而造成的人才流失严重、高学历与复合型人才稀缺、人才选择与培养机制不完善等问题,在研究中通过词频统计法确定胜任特征、问卷调查法验证胜任力模型,并采用灰色关联度法对胜任力模型进行优化得出最终的情报工作胜任力模型。[结果/结论]针对情报工作者胜任力不明确这一根本原因,以优化后的胜任力模型为基础,从个人、企业、政府三个方面对管理咨询行业情报胜任力提升给出科学建议。

基于时空最优低维动力系统的多尺度可压缩湍流数值模拟方法

按照周培源教授关于研究湍流数值模拟建模时必须分析和求解脉动速度场的思想,该研究基于第一性原理,系统地建立了基于时空低维最优动力系统的多尺度可压缩湍流数值模拟方法(LMS方法),并将其应用于多次冲击Richtmyer-Meshkov问题的数值模拟中,首次得到了可压缩湍流的中尺度流场和不同于DNS近似解的湍流近似解.数值结果表明,LMS方法可以用较少的网格获得更精确的湍流近似解.首先解决了研究中遇到的几个问题,为LMS方法的构建铺平了道路.这些问题是:基于湍流的物理特性,提出了湍流大、中、小尺度分解的新概念;找到了box滤波空间相关性的计算方法;指出了湍流建模理论中长期存在的逻辑错误,提出了多尺度湍流模型的概念;讨论了湍流封闭问题的本质和关键,给出了克服湍流封闭问题的数值方法.采用box滤波方法/空间网格平均方法且在大尺度网格的意义下,LMS方法的本质是一种将RANS、LES、DES和DNS等湍流数值模拟方法统一的全新湍流数值模拟方法.需要指出的是,LMS方法也可以作为湍流模型研究的辅助工具,以检验SGS尺度方程/脉动方程中各项所对应的湍流模型是否正确.

柱塞气举排水采气研究现状及展望

气井积液是大部分天然气井在开发中后期存在的主要问题,排水采气工艺是排出井底积液并恢复和提高气井产量的关键技术之一。柱塞气举是一种可以高效排出井底积液的间歇生产工艺,该工艺具有结构简单、成本低等优点,适用于低产井和高气液比井,是目前我国最主要的排水采气技术之一。为此,系统地从柱塞类型及应用场景、柱塞气举速度模型、工作制度优化及故障诊断、新型柱塞气举工艺4个方面进行研究综述,分析了不同模型的优缺点,并总结了8种新型柱塞的特点和应用场景,进而针对目前气井生产效率低和井底积液难排出的情况,展望了柱塞气举工艺的技术发展方向。研究结果表明:(1)相比于国外,中国柱塞气举工艺应用广泛,发展较为成熟,新型柱塞气举工艺的发展进一步扩大了其应用范围。(2)速度模型可精确模拟柱塞在井筒中的运动状态;工作制度优化逐步向智能化发展。(3)目前仍存在柱塞气举工艺智能化程度低和无法适应复杂油气井生产的问题。结论认为,柱塞气举工艺下一步发展应聚焦4个方向:智能化技术的应用、适应复杂井条件下的技术开发、结合其他采气工艺的复合应用、基于气藏—井筒—地面一体化的井群制度优化。该研究成果对进一步推动国内柱塞气举相关研究与应用,具有较好的参考作用,并可以为柱塞气举未来发展方向提供指导。

基于改进蚁群算法的低碳冷链配送路径优化

针对目前冷链配送路径优化中存在的综合配送成本考虑不全面的问题,根据生鲜配送特点,综合考虑基本运输成本、碳排放、制冷、货损以及时间窗约束,建立以综合配送成本最小化为目标的路径优化模型。提出改进蚁群算法对该优化模型进行求解,在初始阶段使用遗传算法生成初期信息素分布,而后使用蚁群算法进行后续寻优搜索,再引入模拟退火算法的Metropolis准则筛选优质解。通过仿真实验验证了优化模型和改进算法的有效性,对低碳可持续发展理念下生鲜品冷链配送路径优化问题研究具有一定意义,助力冷链运输行业向低碳经济转型。

林区智能作业车臂架动力学建模及优化设计

针对林区作业环境复杂等问题,设计一款面向林区作业的林区智能作业车。建立作业车臂架系统的运动学和动力学模型并进行三维软件仿真和优化设计。首先,采用解析几何法与拉格朗日动力学方程结合,建立臂架系统的动力学模型。其次,利用软件NX1899的机构动力学仿真工具Simcenter 3D Motion对臂架系统进行分析,得到臂架系统各油缸驱动力和行程随时间变化曲线。最后,基于响应面BBD(Box-Behnken design)设计响应面试验,对变幅油缸前后两铰点位置进行优化。结果表明,在油缸行程仅增加0.000 04%情况下,油缸驱动力减小2.33%,BBD所提供的试验设计可靠。因此,该动力学模型可为油缸选型和油缸受力优化提供理论依据。

基于奖惩阶梯型碳价机制的能源枢纽低碳优化策略

为进一步降低碳排放水平以及源-荷不确定性对系统运行的影响,文中提出一种基于奖惩阶梯型碳价机制和分布式模型预测控制(distributed model predictive control,DMPC)的能源枢纽(energy hub,EH)日前-日内-实时多时间尺度低碳优化调度策略。引入奖惩阶梯型碳价计算方法,构建EH日前低碳优化调度模型,并制定基于DMPC的日内滚动和实时调整的反馈闭环优化策略,降低源-荷预测误差,提高传统模型预测控制(model predictive control,MPC)的求解效率。在日内阶段,构建以阶梯型碳成本、运行成本和储能调整惩罚成本之和最小为目标的日内滚动优化模型;在实时阶段,分解整体优化问题,建立基于DMPC的多智能体实时调整模型。算例结果表明,文中所提策略能够有效提升系统经济效益,降低源-荷不确定性,实现EH的低碳经济、稳定可靠运行。

县域国土空间用途结构模拟与低碳优化研究——以陕西省渭南市澄城县为例

国土空间用途结构优化是实现城镇低碳绿色高质量发展的重要途径之一。文章以陕西省渭南市澄城县为例,估算并分析2011—2020年的碳排放量,利用Markov-PLUS模型模拟自然发展情景下2025年的国土空间用途结构,并构建多目标规划模型,以经济利益最大化、碳排放最小化、生态效益最大化为目标函数,对所模拟的国土空间用途结构进行优化研究。研究结果表明:1)澄城县碳排放量呈先增后减再增的趋势,近10年碳排放量峰值为84.8170万t。2)模拟自然情况下2025年澄城县碳排放量将达到73.6808万t。3)通过优化,只考虑碳排放效益,对比优化前2025年碳排放量可减少1.0991万t,生态效益增加4954万元;统筹考虑碳排放效益、经济效益与生态效益,对比优化前2025年碳排放量仅减少5578t,但经济效益和生态效益分别增加9098万元和346万元。基于此提出以低碳为目标的国土空间用途结构优化策略。

基于云计算平台的0.4 kV低压带电作业任务调度与资源优化算法的研究

针对电力系统维护中低压带电作业的复杂性和高风险性,提出了一种综合考虑资源利用率和任务完成率的优化模型。该模型以最大化资源利用率和任务完成率为目标,通过引入烟花算法等优化算法,实现了任务的高效调度和资源的合理分配。结果表明,所提算法在提高作业效率的同时,能有效降低带电作业的安全风险,具有良好的实用性和可行性。

基于工业互联网的煤矿综采设备信息模型研究

煤矿综采工作面设备来自不同厂商,接口不统一、数据系统和业务逻辑不同,造成数据壁垒,系统间数据交互慢。基于工业互联网架构,提出了包括设备层、接入层、边缘层、PaaS层、应用层的智能化综采技术架构;基于该架构,将煤矿综采设备作为整体数据对象,设计了一种煤矿综采设备信息模型构建方法,定义了属性、方法、事件、报警4个关键要素和建模规则,从而实现各物理实体与异构系统之间的无缝通信,即对综采设备信息资源进行定义、描述和关联,提供完备、统一的数据对象表达、描述和操作模型;针对综采设备信息模型属性元素过多的问题,提出了一种基于重要度和语义相似度的属性元素优化机制。采用上述方法建立了华阳新材料科技集团有限公司一矿81004工作面综采设备信息模型,对综采设备电动机运行情况进行评价,结果表明基于该模型可监测工作面所有电动机短时间内启动次数、多电动机设备启动功率平衡和运行效率,并形成分析结果,为决策工作提供数据支持。

基于均衡性和能力利用的高速铁路标杆车运行图编制方法研究

高速铁路标杆车的服务品质对于提高铁路客运产品的竞争力具有重要意义,在现场实际编图过程中,高速铁路标杆车须优先铺画,因此标杆车运行图的编制结果决定了整个列车运行图的框架。为了提高高速铁路标杆车的服务品质,考虑标杆车区间运行时分约束、交路接续时间约束等编图基本约束,结合标杆车始发时间约束、成组约束等特殊约束,基于事件活动网络构建以OD服务均衡性和能力利用为目标的整数线性规划模型。设计分阶段迭代求解算法框架,嵌套商业求解器进行求解,第1阶段为不考虑安全间隔时间约束的标杆车分组问题,以OD服务均衡性为目标进行求解;第2阶段以求解得到的标杆车分组为输入,增加安全间隔时间约束,以能力利用为目标进行求解。最后以京广高速铁路为案例,使用实际编图参数,运用优化求解器ILOG CPLEX进行求解,验证模型的有效性和可行性。实验结果表明:优化模型可以在较短时间内高效求解,与实际图对比,上下行OD服务均衡性分别提升38.2%和51.7%,能力损失降低50%,针对不同规模的算例,在200 s内均能得到最优解。针对不同交路接续时间,得出限制标杆车在北京西站的交路接续时间在45 min以内,不会对OD服务均衡性和能力利用造成较大影响。研究结果可为高速铁路标杆车运行图编制及优化提供参考。

基于DRNN-LSTM-IPSO的锅炉经济运行优化目标值

锅炉作为火电厂的重要设备,提高其运行经济性直接影响电厂的生产效益,而锅炉运行参数优化目标值的合理确定是保障锅炉经济运行的关键。首先提出改进的高斯混合模型算法应用于工况划分,即通过划分数据的分散度作为依据并基于马氏距离来构建评价准则函数,以确定聚类数;其次通过构建具备长短时记忆功能的深度循环神经网络(deep recurrent neural network with long-short term memory, DRNN-LSTM)建立各工况区间下的经济模型;最后在经济模型构建的基础上,针对传统粒子群算法容易陷入局部极值问题,通过对惯性权重和加速因子进行调整得到改进的粒子群算法(improved particle swarm optimization, IPSO),可更加精准地在不同工况下进行区间范围内寻优,确定运行参数的优化目标值。实验结果表明,采用本文方法确定的优化目标值对应供电煤耗优于历史最优运行值,说明了该方法在挖掘锅炉优化运行潜力上具有一定的优势,按此方案调整锅炉运行可有效降低能耗水平,以达到锅炉经济运行的目的。

低热水泥与减水剂的相容性研究

本文研究了低热水泥与聚羧酸减水剂和萘系减水剂相容性的差异,并以普通硅酸盐水泥作为基准,通过流动度以及经时损失和流变性能评价了水泥与减水剂的相容性。结果表明:在低热浆体中,萘系减水剂的饱和掺量点大于聚羧酸减水剂;低热水泥展现出不弱于普通硅酸盐水泥的良好相容性,并且低热水泥与聚羧酸减水剂的适应性要好于萘系减水剂;含减水剂的低热水泥净浆的流变特性符合Bingham模型,掺入聚羧酸减水剂的低热水泥浆体呈现为典型的牛顿流体,而掺入萘系减水剂的低热水泥浆体呈现为剪切稀化流体。

基于联邦分割学习与低秩适应的RoBERTa预训练模型微调方法

微调后的大语言模型(Large language models,LLMs)在多任务中表现出色,但集中式训练存在用户隐私泄漏的风险。联邦学习(Federated learning,FL)通过本地训练避免了数据共享,但LLMs庞大的参数量对资源受限的设备和通信带宽构成挑战,导致在边缘网络中部署困难。结合分割学习(Split learning,SL),联邦分割学习可以有效解决这一问题。基于模型深层权重的影响更为显著,以及对部分层的训练准确率略低于整体模型训练的发现,本文按照Transformer层对模型进行分割,同时引入低秩适应(Low⁃rank adaption,LoRA)进一步降低资源开销和提升安全性。因此,在设备端,仅对最后几层进行低秩适应和训练,然后上传至服务器进行聚合。为了降低开销并保证模型性能,本文提出了基于联邦分割学习与LoRA的RoBERTa预训练模型微调方法。通过联合优化边缘设备的计算频率和模型微调的秩,在资源受限的情况下最大化秩,提高模型的准确率。仿真结果显示,仅训练LLMs最后3层的情况下,在一定范围内(1~32)增加秩的取值可以提高模型的准确率。同时,增大模型每轮的容忍时延和设备的能量阈值可以进一步提升模型的准确率。