An approach of improved Multivariate Timing-Random Deep Belief Net modelling for algal bloom prediction. (January 2019)

• Record Type: Journal Article
• Authors: Wang, Li; Zhang, Tianrui; Wang, Xiaoyi; Jin, Xuebo; Xu, Jiping; Yu, Jiabin; Zhang, Huiyan; Zhao, Zhiyao
• Is Part Of: Biosystems engineering. Volume 177(2019)
• Page Start: 130
• Page End: 138
• Publication Date: 2019-01
• Subjects: Algal bloom prediction -- MT-RDBN -- RCRBM -- Bernoulli random number -- Feature extraction; Bioengineering -- Periodicals; Agricultural engineering -- Periodicals; Biological systems -- Periodicals; Génie rural -- Périodiques; Systèmes biologiques -- Périodiques; 631
• DOI: 10.1016/j.biosystemseng.2018.09.005
• Languages: English
• ISSNs: 1537-5110

Abstract

Abstract : Algal bloom formation is a nonlinear time series process for the characterisation factor such as chlorophyll concentration with a variety of interacting influencing factors such as pH, water temperature etc. However, the existing algal bloom prediction methods can't fully reflect complex multi-factor ecological change processes, which result in bloom prediction accuracy being unable to meet requirements. For this problem, multi-factor time series analysis and deep belief net are combined and a Multivariate Timing-Random Deep Belief Net (MT-RDBN) model is proposed. In the MT-RDBN model, the connection between the characterisation factor at current time and the characterisation factor at earlier times, and the connection between the characterisation factor at current time and the influencing factors at both earlier times and current time, are added in input layer. Thus autoregressive model and multivariate regression model of MT-RDBN are constructed. At the same time, the connection between every neuron at current time in hidden layer and every neuron at both earlier times and current time in input layer are added, to realise the description of multi-factor non-linear timing process. In the pre-training phase, multiple Random Conditional Restricted Boltzmann Machines (RCRBM）are constructed by adding Bernoulli random number in front of some parameters, which ensures the randomness of multivariate time series data feature extraction. Then the weight and bias are updated. Finally, back propagation neural network algorithm is used to fine-tune network parameters. The results show that MT-RDBN utilises time series data better and can further improve algal bloom prediction accuracy. Highlights: Cyanobacteria bloom formation is a nonlinear time series ecological process. A multi-factor time series analysis and DBN are combined. We propose a MT-RDBN model.