基于PSO-DNN的平养鸡舍冬季氨气浓度预测模型研究  

Research on prediction model of ammonia concentration in a smart broiler chamber in winter based on PSO-DNN

作　　者：邹修国[1] 宋圆圆 徐泽颖 张世凯 张杰[1] 殷正凌 ZOU Xiuguo;SONG Yuanyuan;XU Zeying;ZHANG Shikai;ZHANG Jie;YIN Zhengling(College of Artificial Intelligence/Jiangsu Key Laboratory of Intelligent Agricultural Equipment,Nanjing Agricultural University,Nanjing 210031,China)

机构地区：[1]南京农业大学人工智能学院/江苏省智能化农业装备重点实验室,江苏南京210031

出　　处：《南京农业大学学报》2021年第1期184-193,共10页Journal of Nanjing Agricultural University

基　　金：中国博士后科学基金资助项目(2015M571782);江苏省农业科技自主创新资金项目[CX(19)2025];中央高校科研业务基本业务费自主创新重点项目(KYTZ201661)

摘　　要：[目的]氨气是鸡舍内影响肉鸡生长发育的主要有害气体,由于冬季鸡舍低通风量会导致氨气浓度超标,使肉鸡的免疫功能下降,导致呼吸系统疾病发生。针对鸡舍氨气预测精度不高、效率不理想等问题,提出基于粒子群算法(particle swarm optimization,PSO)优化深度神经网络(deep neural network,DNN)的预测模型,实现冬季氨气浓度预警并及时调控鸡舍内氨气的浓度。[方法]选取自建平养鸡舍环境参数数据(温度、相对湿度和氨气浓度)和鸡自身情况数据(鸡龄和鸡进入鸡舍时间)建立模型,对鸡舍内未来1 h氨气浓度进行预测。PSO-DNN预测模型首先采用PSO优化DNN中的batch_size参数,以平均绝对误差(mean absolute error,MAE)作为目标函数,经过多次迭代后,得到最佳的batch_size,再以此构建DNN模型,以数据集的前70%数据作为训练集进行DNN模型训练,经过DNN的线性运算和激活运算后,采用数据集的后30%数据对模型进行验证,并对模型进行评估。[结果]将PSO-DNN模型与DNN和随机森林模型对比,PSO-DNN模型氨气预测结果的MAE为1.886 mg·m^-3,DNN和随机森林模型预测的MAE分别为4.297和2.855 mg·m^-3。[结论]PSO-DNN模型的预测精度最高,与DNN和随机森林模型预测结果相比,其MAE分别降低56.1%和33.9%,可为平养鸡舍内氨气浓度预测提供方法参考,有助于及时、准确地调控鸡舍内氨气浓度。[Objectives]Ammonia is the main harmful gas that affects the growth and development of broilers.Due to the low ventilation rate in winter,the ammonia concentration will exceed the standard,which will reduce the immune function of broilers and lead to respiratory diseases.Aiming at the problems of low accuracy and low efficiency of ammonia prediction in broiler chambers,a prediction model based on particle swarm optimization(PSO)optimized deep neural network(DNN)was proposed.It can realize early warning of ammonia concentration and regulate the quality of the ammonia environment in broiler chambers.[Methods]This research selected the environmental parameter data(temperature,relative humidity,and ammonia concentration)of self-built smart broiler chamber and the data of the broilers’conditions(broilers age and time when the broilers entered the broiler chamber)to build a model,so as to predict the ammonia concentration in the broiler chamber in the next hour.The PSO-DNN prediction model first used PSO to optimize the batch_size parameter in DNN and used the minimum average absolute error as the objective function.After 100 iterations,the best batch_size was obtained,and then the DNN model was constructed.The PSO-DNN model was trained with the first 70%of the data set as the training set.After the linear calculation and activation calculation of DNN,the last 30%of the data set was used to verify the model and evaluate the model.[Results]The average absolute error of the ammonia prediction result of the PSO-DNN model was 1.886 mg·m^-3.Compared it with the DNN and the random forest models,the average absolute errors of the DNN and random forest models were 4.297 and 2.855 mg·m^-3.[Conclusions]The results show that the PSO-DNN model has the highest prediction accuracy.Compared with the prediction results of the DNN and random forest models,the average absolute error is reduced by 56.1%and 33.9%,which can provide technical methods for the prediction of ammonia concentration in broiler chambers for reference.It will h

关 键 词：平养鸡舍 氨气浓度 深度神经网络 粒子群算法 随机森林 

分 类 号：S831.4[农业科学—畜牧学]