领域知识引导的作物模型参数智能优化框架研究

发布时间：2018-06-25 13:20

本文选题：作物生长模型 + 遗传算法　；参考：《南京农业大学》2015年硕士论文

【摘要】：作物模型是以作物生长发育机理为基础,对作物生理过程与环境和技术的关系加以理论概括和量化分析的数学模型,已经在农作物估产、农田管理决策等领域广泛应用。作物模型在应用过程中需要针对不同环境条件重新优化其品种参数。遗传算法作为一种高效的启发式搜索技术,已在作物生长模型参数优化问题中得到了良好的应用。但存在以下问题:由于作物模型本身结构复杂、实测数据误差、参数智能优化过程环节众多,给作物模型参数智能优化带来了很大的不确定性;由于遗传算法的随机搜索机制,会导致优化过程中出现目标拟合较好但不符合生理学特性的优化结果;随着模型的广泛应用,模型不确定性评价、参数智能优化等工作需要借助计算机软件工具进行快速实现,但目前的软件工具功能单一,难以面向通用的作物模型领域广泛应用。针对上述问题,论文主要贡献包括:(1)分析基于协同进化遗传算法的水稻生长模型参数优化框架(BGA-CMPOF)的不确定性。BGA-CMPOF中的不确定性来源包括:实测数据误差、适应度函数设计、参数优化策略、优化算法性能等环节,本文设计了评价BGA-CMPOF框架不确定性的策略,从目标变量选择、适应度权重设置、分阶段参数优化策略以及不同优化算法等角度,分析了 BGA-CMPOF在RiceGrow水稻生长模型参数优化中的不确定性。以汕优63在徐州等地的实测数据进行试验,结果表明:1)BGA-CMPOF框架可有效优化RiceGrow模型品种参数,各参数优化的相对误差在7%以内;2)选取LAI、各器官(茎、叶、穗)生物量作为目标变量的效果较好,穗生物量、总生物量、LAI的NRMSE分别减小了 0.32%、1.52%和 1.73%,RMSE 分别减小了 8kg/ha、123.1 kg/ha、0.08;MAD分别减小了 5.44 kg/ha、105.1 kg/ha、0.07;3)分阶段参数优化策略对于BGA-CMPOF而言效果不明显,各目标NRMSE的差距1%;4)MECA算法比IAGA算法的优化精度略高,但耗时较长,不利于模型参数的快速优化。(2)提出基于领域知识引导遗传算法的作物生育期模型参数优化方法。建立作物模型参数智能优化领域知识库,通过约束模型初始参数范围、确定调参关键物候期、扩展物候期实测值、提炼方向算子等环节,对基于遗传算法的作物模型参数优化框架进行约束和引导。本文以WheatGrow小麦生育期模型为应用对象,针对济南13号小麦品种在徐州、济宁、潍坊、临沂等地的实测数据的参数优化试验结果表明:1)四个地点的初步试验验证结果的RMSE分别达到了:1.51d、2.05d、0.72d和1.08d,R2均0.99,MAD分别为1.1d、1.6d、0.55d和0.8d,模拟效果较好,但部分参数存在不符合生物学特性的参数值;2)通过约束参数PVT的初始范围[30,40],加入品种参数范围约束后,各地点的调参结果的各项指标均有小幅增加,RMSE分别为 1.67d、2.12d、1.09d和 1.58d,R2均0.99,MAD 分别为 1.4d、1.6d、0.8d和 1.3d,但PVT的参数值均符合济南13号半冬性的品种特性;3)扩展关键物候期的实测值后,四地的验证结果RMSE分别为1.24d、1.56d、1.22d和1.48d,R2分别为0.997、0.993、0.997和0.991,MAD分别为1d、1.2d、1.8d和1.2d,并且品种参数IE符合生物学特性,结果表明,扩展调参物候期数据进行参数优化,在保证优化效果的同时,能够起到约束品种参数的作用。4)加入方向算子后,IAGA算法的收敛代数分别减少了 8代和3代,方向算子能够加快算法的收敛速度。(3)研制作物模型参数优化及不确定性分析工具(CMPOAT)采用构件化软件中基于框架的软件开发方法,开发基于动态组装框架的作物模型自动调参及不确定性分析工具,能够根据用户需求,实现作物模型、进化算法、数据处理等业务的动态组装。系统实现了:数据管理与分析、作物模型分析、调参计算、专家知识库管理、组件库管理等功能。以WheatGrow模型和IAGA算法为对象的应用案例表明,CMPOAT能够分析作物模型不确定性并进行模型参数优化,为作物模型的分析和应用提供了有力的软件工具。
[Abstract]:The crop model is a mathematical model based on the mechanism of crop growth and development and the theoretical generalization and quantitative analysis of the relationship between crop physiological process and environment and technology. It has been widely used in the fields of crop yield estimation and farmland management decision. As an efficient heuristic search technique, genetic algorithm has been well applied in the parameter optimization of crop growth model. However, the following problems are as follows: because the structure of the crop model is complex, the measured data error and the parameter intelligent optimization process are numerous, it has brought a great deal to the intelligent optimization of the crop model parameters. As a result of the stochastic search mechanism of genetic algorithm, the optimization results will lead to the optimization results of better target fitting but not in conformity with the physiological characteristics. With the extensive application of the model, the model uncertainty evaluation and parameter intelligent optimization need to be implemented quickly with the aid of computer software tools, but the current software tools are used. The main contributions of this paper are as follows: (1) the uncertainty sources in the uncertainty.BGA-CMPOF of the parameter optimization framework of the rice growth model (BGA-CMPOF) based on Coevolutionary Genetic Algorithm (coevolutionary genetic algorithm) include the measured data error, the fitness function design, and the parameter design. In this paper, the strategy of evaluating the uncertainty of BGA-CMPOF framework is designed in this paper. The uncertainty of BGA-CMPOF in the parameter optimization of RiceGrow rice growth model is analyzed from the selection of the target variables, the setting of fitness weight, the optimization strategy of the phased parameters and the different optimization algorithms. Experiments were conducted in Xuzhou and other places. The results showed that: 1) the BGA-CMPOF framework could effectively optimize the parameter of RiceGrow model, and the relative error of each parameter was less than 7%; 2) LAI was selected as the target variable, and the ear biomass, total biomass, and NRMSE of LAI were reduced by 0.32%, 1.52% respectively. And 1.73%, RMSE reduced 8kg/ha, 123.1 kg/ha, 0.08; MAD decreased 5.44 kg/ha, 105.1 kg/ha, 0.07; 3). The phase parameter optimization strategy is not obvious to BGA-CMPOF, the gap 1% of each target NRMSE, 4) MECA algorithm is slightly higher than IAGA algorithm, but it takes longer time and is not conducive to the rapid optimization of model parameters. (2) The parameter optimization method of crop growth period model based on domain knowledge guided genetic algorithm is proposed. The knowledge base of intelligent optimization field of crop model parameters is set up. By restricting the range of the initial parameters of the model, the key phenology period of the adjustment parameter, the measured value of the phenology period and the direction operator are refined, and the parameter of the crop model based on the genetic algorithm is obtained. The optimization of the framework for constraints and guidance. This paper takes the WheatGrow wheat growth period model as the application object. The results of the parameters optimization test of the measured data of Ji'nan No. 13 wheat variety in Xuzhou, Jining, Weifang, Linyi and other places show that: 1) the preliminary test of four locations verified that the results of the results were as follows: 1.51d, 2.05d, 0.72d and 1.08d, R2 0 respectively. .99, MAD are 1.1d, 1.6d, 0.55d and 0.8d, and the simulation results are better, but some parameters have the parameter values that do not conform to the biological characteristics; 2) after the restriction of the initial range [30,40] of the parameters of the parameter PVT, the parameters of the parameters of the parameter range of the local points are increased slightly, and RMSE is 1.67d, 2.12d, 1.09d, and, respectively. All 0.99, MAD are 1.4d, 1.6d, 0.8d and 1.3d respectively, but the parameters of PVT are in line with Ji'nan 13 and half winter variety characteristics; 3) after expanding the measured values of the key phenology, the verification results of the four regions are 1.24d, 1.56D, 1.22d and 1.48d, respectively, and 0.991, respectively. According to the biological characteristics, the results show that the parameter optimization of the phenology period data is extended, while the effect of the optimization is guaranteed, while the effect of the constrained variety parameters can be played.4), the convergence algebra of the IAGA algorithm is reduced by 8 and 3 generations respectively. The direction operator can speed up the convergence speed of the algorithm. (3) to develop the parameter of the crop model. Optimization and uncertainty analysis tool (CMPOAT) adopts a framework based software development method in component-based software, and develops a tool for automatic parameter adjustment and uncertainty analysis of crop models based on dynamic assembly framework. It can implement the dynamic assembly of crop model, evolutionary algorithm and data processing based on user requirements. According to management and analysis, crop model analysis, parameter calculation, expert knowledge base management, component library management and other functions. The application cases of WheatGrow model and IAGA algorithm show that CMPOAT can analyze crop model uncertainty and optimize model parameters, and provide a powerful software tool for crop model analysis and application.
【学位授予单位】：南京农业大学
【学位级别】：硕士
【学位授予年份】：2015
【分类号】：S126;TP18

【参考文献】