An approach of recursive timing deep belief network for algal bloom forecasting

• Published in: Neural computing & applications Vol. 32; no. 1; pp. 163 - 171
• Main Authors: Wang, Li, Zhang, Tianrui, Jin, Xuebo, Xu, Jiping, Wang, Xiaoyi, Zhang, Huiyan, Yu, Jiabin, Sun, Qian, Zhao, Zhiyao, Xie, Yuxin
• Format: Journal Article
• Language: English
• Published: London Springer London 2020; Springer Nature B.V
• Subjects: Algae; Algorithms; Artificial Intelligence; Belief networks; Brain- Inspired computing and Machine learning for Brain Health; Computational Biology/Bioinformatics; Computational Science and Engineering; Computer Science; Data Mining and Knowledge Discovery; Forecasting; Image Processing and Computer Vision; Mathematical models; Neural networks; Nonlinear dynamics; Probability and Statistics in Computer Science; Recursive methods; Time series; Dynamic nonlinear process; Algal bloom; Recursive time series deep belief network; Forecasting
• ISSN: 0941-0643; 1433-3058
• DOI: 10.1007/s00521-018-3790-9

The forecasting methods of water bloom in existence are hard to reflect nonlinear dynamic change in algal bloom formation mechanism, leading to poor forecasting accuracy of bloom. To solve this problem, this paper deeply analyzes the generation process of algal bloom, introduces the recursive time series algorithm into the deep belief network model and improves the model structure and training algorithm, and proposes a forecasting method based on the recursive timed deep belief network model. The model introduces the current moments and historical time values of the characterization factors and influencing factors at the input layer, and increases the connection between the input layer and the hidden layer of the deep belief network. A recursive algorithm is used to establish the relationship between the current time value of the characterization factor and the historical time value of the characterization factor, and the connection between the current time value of the hidden layer and the influencing factor is increased. By re-extracting the characteristics of the hidden layer at each moment, and then fine tuning the network parameters by the BP neural network, a recursive timing deep belief network model is finally constructed. The results show that compared with the existing forecasting methods, this method can extract the characteristics of time series data more accurately and completely to deal with the dynamic nonlinear process and can further improve the forecast accuracy of algal blooms.