基于CNN-GRU混合模型的养殖工船水体溶解氧预测研究

苏辉锋; 丁乐声; 王绪旺; 陈木生; 陈潇

doi:10.12131/20220298

基于CNN-GRU混合模型的养殖工船水体溶解氧预测研究

doi: 10.12131/20220298

1.
南方海洋科学与工程广东省实验室 (湛江)，广东湛江 524000
2.
大连理工大学宁波研究院，浙江宁波 315700

基金项目: 广东省海洋经济发展 (海洋六大产业) 专项资金资助项目 (GDNRC[2021]42)；湛江市海洋装备和海洋生物揭榜挂帅制人才团队专项资金资助项目 (2021E05034)；南方海洋科学与工程广东省实验室 (湛江) 项目 (ZJW-2019-01)

详细信息

作者简介:
苏辉锋 (1995—)，男，研究实习员，硕士，研究方向为海洋工程装备、渔业设备研发。E-mail: suhuifeng@zjblab.com

通讯作者:
陈　潇 (1986—)，男，高级工程师，本科，研究方向为海洋工程装备、渔业设备研发。E-mail: chenxiao_ship@163.com

中图分类号: S 967.9
计量
- 文章访问数: 289
- HTML全文浏览量: 71
- PDF下载量: 31
- 被引次数: 0
出版历程
- 收稿日期: 2022-11-23
- 修回日期: 2023-03-13
- 录用日期: 2023-03-31
- 网络出版日期: 2023-04-28
- 刊出日期: 2023-08-05

Prediction of dissolved oxygen in water of aquaculture ship based on CNN-GRU hybrid model

1.
Southern Marine Science and Engineering Guangdong Laboratory (Zhanjiang), Zhanjiang 524000, China
2.
Ningbo Institute of Dalian University of Technology, Ningbo 315700, China

摘要

摘要: 溶解氧 (Dissolved oxygen, DO) 是影响养殖工船水产品健康生长的重要因素，准确预测DO对提高水产品产量和品质具有重要意义。为提高DO预测精度，以卵形鲳鲹 (Trachinotus ovatus) 养殖试验采集的数据为样本，使用卷积神经网络 (Convolutional neural network, CNN) 和门控循环单元 (Gated recurrent unit, GRU) 方法建立养殖工船水体DO预测混合模型，通过Pearsons相关性分析，选用DO、温度、pH和循环水流量4个预测因子进行训练和校准，预测了DO含量。通过与CNN、GRU和长短期记忆 (Long short-term memory, LSTM) 模型进行对比，所建模型在各项评价指标中的性能均最优，其均方根误差 (Root mean square error, RMSE)、平均绝对误差 (Mean absolute error, MAE) 和决定系数R²分别为0.119、0.084和0.976。结果表明，所建模型的预测精度最高，可以满足养殖工船实际生产中对DO预测的需求，为养殖工船生产过程中DO的监控和预警提供参考。
- 养殖工船 /
- 溶解氧 /
- 卷积神经网络 /
- 门控循环单元
Abstract: Dissolved oxygen (DO) content is a critical factor that affects the healthy growth of aquatic products in aquaculture ships. Accurate prediction of DO content is necessary to improve aquatic production and quality. To increase the accuracy of DO prediction , based on the data collected from a Trachinotus ovatus culture experiment, we established a hybrid model for DO prediction in aquaculture ships by applying the convolutional neural network (CNN) and gated recurrent unit (GRU) methods. Based on Pearson correlation analysis, we selected four predictors, namely dissolved oxygen content, temperature, pH value and circulating water flow, which were trained and calibrated to predict the DO content. The model proposed in this paper outperformed CNN, GRU and long short-term memory (LSTM) models in all evaluation indexes, and its root mean square error (RMSE), mean absolute error (MAE) and determination coefficient R² were 0.119, 0.084 and 0.976, respectively. The results indicate that the model proposed in this paper has the greatest prediction precision and can meet the demand for DO content prediction in actual production of aquaculture ships, which provides references for monitoring and early warning of DO content in the production process of aquaculture ships.
- Aquaculture ship /
- Dissolved oxygen /
- Convolutional neural network /
- Gated recurrent unit

HTML全文

图 1 养殖舱

Figure 1. Aquaculture water tank

下载: 全尺寸图片幻灯片

图 2 养殖舱监控系统结构简图和实物图

Figure 2. Schematic diagram and physical diagram of aquaculture tank monitoring system

下载: 全尺寸图片幻灯片

图 3 LSTM 单元示意图

Figure 3. Schematic diagram of LSTM unit

下载: 全尺寸图片幻灯片

图 4 GRU 单元示意图

Figure 4. Schematic diagram of GRU unit

下载: 全尺寸图片幻灯片

图 5 CNN-GRU 网络结构图

Figure 5. Structure diagram of CNN-GRU network

下载: 全尺寸图片幻灯片

图 6 4 种模型预测结果

Figure 6. Prediction results of four models

下载: 全尺寸图片幻灯片

表 1 水质传感器规格

Table 1. Water quality sensor specification

水质参数 Parameter of water quality	测量范围 Range of measurement	精度 Precision	分辨率 Resolution ratio	型号 Model	品牌 Brand
溶解氧质量浓度 DO/(mg·L⁻¹)	0~20	±2% (满量程百分比)		LDO II	哈希 Hach
温度 Temperature/℃	0~50	±0.5	0.1 ℃	MPS-400	凯米斯 Chemins
pH	0~14	±0.1	0.01	MPS-400	凯米斯 Chemins
盐度 Salinity/‰	0~100	±3.5% (满量程百分比)	0.1‰	MPS-400	凯米斯 Chemins

下载: 导出CSV

表 2 4 种预测模型的超参数设置

Table 2. Hyperparameter setting of four prediction models

模型 Model	参数 Paramenter
卷积神经网络 CNN	卷积核个数：3 卷积核大小：32
长短期记忆 LSTM	隐藏层个数：128 全连接层神经元个数：128
门控循环单元 GRU	隐藏层个数：128 全连接层神经元个数：128
卷积神经网络-门控循环单元 CNN-GRU	卷积核个数：3 卷积核大小：32 GRU隐藏层个数：128 全连接层神经元个数：128

下载: 导出CSV

表 3 4 种模型的预测性能

Table 3. Predictive performance of four models

模型 Model	均方根误差 RMSE/ (mg·L⁻¹)	平均绝对误差 MAE/ (mg·L⁻¹)	决定性系数 R²
卷积神经网络 CNN	0.173	0.132	0.945
长短期记忆 LSTM	0.143	0.120	0.956
门控循环单元 GRU	0.138	0.114	0.966
卷积神经网络-门控循环单元 CNN-GRU	0.119	0.084	0.976

下载: 导出CSV

参考文献(33)

[1]	刘敏. 大型养殖工船运营实践与展望[J]. 船舶工程, 2021, 43(4): 18-23.
[2]	崔铭超, 鲍旭腾, 王庆伟. 我国深远海养殖设施装备发展研究[J]. 船舶工程, 2021, 43(4): 31-38.
[3]	AHMED N, TURCHINI G M. Recirculating aquaculture systems (RAS): environmental solution and climate change adaptation[J]. J Clean, 2021, 297: 126604. doi: 10.1016/j.jclepro.2021.126604
[4]	SAMARAS A, TSOUKALI P, KATSIKA L, et al. Chronic impact of exposure to low dissolved oxygen on the physiology of Dicentrarchus labrax and Sparus aurata and its effects on the acute stress response[J]. Aquaculture, 2023, 562: 738830. doi: 10.1016/j.aquaculture.2022.738830
[5]	JIANG X Y, DONG S L, LIU R X. Effects of temperature, dissolved oxygen, and their interaction on the growth performance and condition of rainbow trout (Oncorhynchus mykiss)[J]. J Therm Biol, 2021, 98: 102928. doi: 10.1016/j.jtherbio.2021.102928
[6]	NEILAN R M, ROSE K. Simulating the effects of fluctuating dissolved oxygen on growth, reproduction, and survival of fish and shrimp[J]. J Theor Biol, 2014, 343: 54-68. doi: 10.1016/j.jtbi.2013.11.004
[7]	LIU Y Q, ZHANG Q, SONG L H, et al. Attention-based recurrent neural networks for accurate short-term and long-term dissolved oxygen prediction[J]. Comput Electron Agric, 2019, 165: 104964. doi: 10.1016/j.compag.2019.104964
[8]	杨争光, 范良忠. 基于MEC-BP神经网络在水产养殖水质预测中的应用[J]. 计算机与现代化, 2015(6): 32-36. doi: 10.3969/j.issn.1006-2475.2015.06.007
[9]	汤铠源. 基于 HDE 算法的 FA-BP 神经网络溶解氧浓度预测控制研究[D]. 马鞍山: 安徽工业大学, 2019: 48-55.
[10]	YANG Y T, TAI H J, LI D L. Real-time optimized prediction model for dissolved oxygen in crab aquaculture ponds using back propagation neural network[J]. Sens Lett, 2014, 12(3/4): 723-729.
[11]	袁红春, 潘金晶. 改进递归最小二乘RBF神经网络溶解氧预测[J]. 传感器与微系统, 2016, 35(10): 20-23. doi: 10.13873/J.1000-9787(2016)10-0020-04
[12]	刘晨, 李莎, 丛孙丽, 等. 基于EEMD和萤火虫算法优化SVM的溶解氧预测[J]. 计算机仿真, 2021, 38(1): 359-365.
[13]	施珮, 匡亮, 袁永明, 等. 基于改进极限学习机的水体溶解氧预测方法[J]. 农业工程学报, 2020, 36(19): 225-232. doi: 10.11975/j.issn.1002-6819.2020.19.026
[14]	GUO J J, DONG J Q, ZHOU B, et al. A hybrid model for the prediction of dissolved oxygen in seabass farming[J]. Comput Electron Agric, 2022, 198: 106971. doi: 10.1016/j.compag.2022.106971
[15]	孙龙清, 吴雨寒, 孙希蓓, 等. 基于IBAS和LSTM网络的池塘水溶解氧含量预测[J]. 农业机械学报, 2021, 52(S1): 252-260.
[16]	ZHANG K Y, CHEN J L, ZHANG T C, et al. A compact convolutional neural network augmented with multiscale feature extraction of acquired monitoring data for mechanical intelligent fault diagnosis[J]. J Manuf Syst, 2020, 55: 273-284. doi: 10.1016/j.jmsy.2020.04.016
[17]	BHANDARI H N, RIMAL B, POKHREL N R, et al. Predicting stock market index using LSTM[J]. Mach Learn Appl, 2022, 9: 100320.
[18]	庞传军, 张波, 余建明. 基于LSTM循环神经网络的短期电力负荷预测[J]. 电力工程技术, 2021, 40(1): 175-180.
[19]	ZARZYCKI K, ŁAWRYŃCZUK M. Advanced predictive control for GRU and LSTM networks[J]. Inf Sci, 2022, 616: 229-254. doi: 10.1016/j.ins.2022.10.078
[20]	AYZEL G, HEISTERMANN M. The effect of calibration data length on the performance of a conceptual hydrological model versus LSTM and GRU: a case study for six basins from the CAMELS dataset[J]. Comput Geosci, 2021, 149: 104708. doi: 10.1016/j.cageo.2021.104708
[21]	OZDEMIR A C, BULUŞ K, ZOR K. Medium- to long-term nickel price forecasting using LSTM and GRU networks[J]. Resour Policy, 2022, 78: 102906. doi: 10.1016/j.resourpol.2022.102906
[22]	DEY P, CHAULYA S K, KUMAR S. Hybrid CNN-LSTM and IoT-based coal mine hazards monitoring and prediction system[J]. Process Saf Environ Prot, 2021, 152: 249-263. doi: 10.1016/j.psep.2021.06.005
[23]	ATILA O, ŞENGÜR A. Attention guided 3D CNN-LSTM model for accurate speech based emotion recognition[J]. Appl Acoust, 2021, 182: 108260. doi: 10.1016/j.apacoust.2021.108260
[24]	REN J, YU Z P, GAO G L, et al. A CNN-LSTM-LightGBM based short-term wind power prediction method based on attention mechanism[J]. Energy Rep, 2022, 8: 437-443.
[25]	ARASH M, HAMID T, BEHNAM M L, et al. Hybrid CNN-LSTM approaches for identification of type and locations of transmission line faults[J]. Int J Electr Power Energy Syst, 2022, 135: 107563. doi: 10.1016/j.ijepes.2021.107563
[26]	贾睿, 杨国华, 郑豪丰, 等. 基于自适应权重的CNN-LSTM&GRU组合风电功率预测方法[J]. 中国电力, 2022, 55(5): 47-56.
[27]	高震宇. 基于深度卷积神经网络的图像分类方法研究及应用[D]. 合肥: 中国科学技术大学, 2018: 4-5.
[28]	王锦涛, 文晓涛, 何易龙, 等. 基于CNN-GRU神经网络的测井曲线预测方法[J]. 石油物探, 2022, 61(2): 276-285.
[29]	ARUNKUMAR K E, KALAGA D V, KUMAR C M S, et al. Comparative analysis of Gated Recurrent Units (GRU), long short-term memory (LSTM) cells, autoregressive integrated moving average (ARIMA), seasonal autoregressive integrated moving average (SARIMA) for forecasting COVID-19 trends[J]. Alex Eng J, 2022, 61(10): 7585-7603. doi: 10.1016/j.aej.2022.01.011
[30]	CHUNG J, GULCEHRE C, CHO K H, et al. Empirical evaluation of gated recurrent neural networks on sequence modeling[J]. Arxiv Org, 2014. DOI: 10.48550/arXiv.1412.3555.
[31]	GAO S, HUANG Y F, ZHANG S, et al. Short-term runoff prediction with GRU and LSTM networks without requiring time step optimization during sample generation[J]. J Hydrol, 2020, 589: 125188. doi: 10.1016/j.jhydrol.2020.125188
[32]	CAMASTRA F, CAPONE V, CIARAMELLA A, et al. Prediction of environmental missing data time series by Support Vector Machine Regression and Correlation Dimension estimation[J]. Environ Model Softw, 2022, 150: 105343. doi: 10.1016/j.envsoft.2022.105343
[33]	CAO X K, REN N, TIAN G L, et al. A three-dimensional prediction method of dissolved oxygen in pond culture based on Attention-GRU-GBRT[J]. Comput Electron Agric, 2021, 181: 105955. doi: 10.1016/j.compag.2020.105955