数据驱动的集约化水健康禽养殖环境多参数非线性集成预测模型研究

The environmental deterioration of intensive waterfowl breeding is an important factor leading to the loss of fine varieties, degeneration of laying performance, induced disease outbreaks, and even large-scale deaths. Accurate prediction of breeding environment has always been a thorny problem that needs to be solved in the waterfowl breeding industry. In this project, key parameters such as ammonia, CO2, light, and temperature are used as research objects. Temporal and spatial variability functions and multivariate analysis of variance are used to study the temporal and spatial variability of environmental parameters under different conditions, and the breeding environment and egg production performance and the cultivation of improved varieties are explored. Interaction relationship; Using new multi-dimensional signal processing technology to find suitable method for extracting environmental parameters of waterfowl breeding, constructing high-quality environmental parameter characteristic matrix; Researching big data mining, ensemble learning, integrator diversity selection criteria, swarm intelligent multi-target optimize the algorithm, explore the method of self-adaptive determination of nonlinear combined weights suitable for aquaculture environment, preliminarily establish an integrated predictive model, Combination weight coefficient is realized by using swarm intelligent multi-objective optimization algorithm and adaptive optimization model parameter adjustment, has built the learning of data driven, high precision and dynamic nonlinear integrated forecast model of intensive water fowl breeding environment, precision prediction of intensive water fowl breeding environment, early warning and control, for intensive water fowl breeding environment strategy management , breeding of improved varieties, improvement of egg production performance and prevention of disease outbreaks to provide scientific basis and method guidance.

集约化水禽养殖环境恶化是导致优良品种退化、产蛋性能降低、诱导疾病爆发甚至大批量死亡的重要因素，而养殖环境精准预测一直是水禽养殖领域亟需解决的棘手难题。本项目以氨气、温度、光照和CO2等关键参数为研究对象，采用时空变异函数和多因素方差分析方法，研究不同条件下环境参数时空变异特性，挖掘养殖环境与产蛋性能、优良品种培育相互作用关系；利用新型多维信号处理技术，找到适宜于水禽养殖环境参数特征提取方法，构建高质量的环境参数特征矩阵；研究大数据挖掘、集成学习、集成器多样性选择准则、群集智能多目标优化算法，探索适宜于养殖环境非线性组合权重自适应确定方法，初步建立集成预测模型；利用群集智能多目标优化算法自适应优化调整组合权重系数和模型参数，构建自学习、精度高、数据驱动的水禽养殖环境非线性集成预测模型，实现养殖环境精准预测。为水禽养殖环境精准调控、优良品种培育、产蛋性能提升、预防疾病爆发等提供科学依据与方法。

为解决集约化水禽养殖环境精准预测难题，本课题开展了数据驱动的集约化水禽健康养殖环境多参数非线性集成预测模型研究。提出了基于变分模态分解、基于经验模式分解、完全自适应噪声集合经验模态分解、过改进核主成分分析、灰色关联分析、多粒度粗糙集、多元回归分析、经验正交函数等理论与方法、系统动力学和主成分分析法等养殖环境时空变异性，探索了养殖环境参数对水禽健康生长繁育生物特性的影响及相关性分析；提出了基于小波变换、集合经验模态分解(EEMD)和深度卷积神经网络(deep CNN)等多维信号处理技术的水禽养殖环境特征快速精准提取，获得多尺度的水禽养殖环境参数特征矩阵；采用改进粒子群算法（PSO）、麻雀搜索算法（SSA）、改进灰狼算法研究（GWO）等群集智能优化算法，解决水禽养殖环境的非线性集成权重系数和模型多参数优化调整问题；研究了BP神经网络(BPNN)、自回归滑动平均模型（ARMA）、极限学习机（ELM）、门控循环神经网络（GRU）等单项预测方法性能和适用范围，集成预测有效性评价准则，提出了提出了基于VMD-DeepAR-SSA的水禽养殖环境NH3区间组合预测模型、提出了基于CEEMDAN-LZC-GOBLPSO-GRU的水禽养殖环境温度组合预测模型、提出了基于PCA-SVR-ARMA的狮头鹅养殖禽舍气温组合预测模型、基于SRU序列到序列模型的水禽养殖环境湿度预测、提出了基于XGBoost的水禽养殖粉尘预测研究、提出了基于平行向量自回归时间序列VAR（2）的蛋鸭室外养殖环境预测模型等多种水禽养殖环境非现金集成预测模型，实现提高水禽养殖环境预测精度，本课题研究为水禽养殖环境非线性集成预测和水禽健康养殖提供了科学方法和理论依据。依托本项目资助，共发表论文34篇，其中SCI论文15篇，EI论文5篇，中文核心4篇；申请专利7件、其中授权发明专利4件，实用新型专利授权1件，授权计算机软件著作权9项，获广东省计算机学会科学技术奖一等奖、广东省农业技术推广奖二等奖等9项。培养研究生9人，参加人员职称晋升8人次。