| 摘要: |
| 为寻求准确的鸡舍氨气浓度预测方法，构建基于双阶段注意力机制和长短时记忆神经网络(Long short-term memory，LSTM)的鸡舍氨气浓度预测模型，将该模型应用于山东省商河县某蛋鸡养殖场，采集二氧化碳(CO2)、氧气(O2)和氨气(NH3)的体积分数，细颗粒物(PM2.5)质量浓度，温度，相对湿度时间序列数据对模型进行验证，并与支持向量回归(Support vector regression，SVR)、人工神经网络(Artificial neural network，ANN)模型和无注意力机制的LSTM模型对比研究。结果表明:1)不同时间窗口T下NH3体积分数预测精度不同。T∈{2，3，4，8}时，均方根误差(Root mean square error，RMSE)分别为0.433 4，0.394 8，0.379 9和0.405 1 μL/L，平均绝对误差(Mean absolute error，MAE)分别为0.267 4，0.262 9，0.228 9和0.272 4 μL/L；2)基于双阶段注意力机制和LSTM的鸡舍NH3浓度预测模型在RMSE和MAE评价指标框架下优于SVR、ANN和无注意力机制的LSTM模型。基于双阶段注意机制和LSTM的模型能较好地对鸡舍氨气浓度进行预测，可为鸡舍氨气浓度预测及调控提供技术支持。 |
| 关键词: 注意力机制 长短时记忆神经网络 循环神经网络 编码器-解码器 鸡舍 氨气浓度 |
| DOI：10.11841/j.issn.1007-4333.2021.06.19 |
| 投稿时间：2020-09-07 |
| 基金项目:山东省家禽疫病诊断与免疫重点实验室开放课题基金(SDPDI201806) |
|
| Ammonia concentration forecasting algorithm in layer house based on two-stage attention mechanism and LSTM |
|
GUO Xiangyun1,LIAN Jinghua2,LI Huimin2,SUN Kai2
|
| (1.School of Information Management, Beijing Information Science and Technology University, Beijing 100192, China;2.Poultry Institute, Shandong Academy of Agricultural Science, Jinan 250023, China) |
| Abstract: |
| To precisely forecast NH3 concentration in layer house, this study proposes a NH3 concentration prediction model based on two-stage attention mechanism and long short-term memory(LSTM). In order to validate the model, time series of carbon dioxide(CO2), oxygen(O2)and ammonia(NH3)volume fraction, particulate matter(PM2. 5)mass concentration, temperature and relative humidity in layer house in Shanghe County Shandong Province are collected and input to two-stage attention mechanism and LSTM base model. The results show that the model could predict NH3 concentration in layer house accurately. Under different window size T∈{2, 3, 4, 8}, the root mean square error(RMSE)are 0. 433 4, 0. 394 8, 0. 379 9 and 0. 405 1 μL/L, respectively; The mean absolute error(MAE)are 0. 267 4, 0. 262 9, 0. 228 9 and 0. 272 4 μL/L, respectively. The model performance in this research is better than the support vector regression(SVR), artificial neural network(ANN)and common LSTM without attention mechanism by RMSE and MAE showing that the model has great potential to be used in NH3 concentration prediction and control in layer house. |
| Key words: attention mechanism long short-term memory neural network recurrent neural network encoder-decoder layer house ammonia concentration |
