Integrating multi-modal data into AFSA-LSTM model for real-time oil production prediction

Oil production prediction plays an important role in the development adjustment and optimization. Most of the existing works solve this problem by identifying the impact of historical production conditions on production via sequential analysis. Although these works have better predicting accuracy compared with traditional techniques, they still face two limitations: (i) data from a single modal cannot provide comprehensive information for prediction models; and (ii) the hyper-parameters of deep neural networks are usually set manually, which cannot guarantee the optimality. To address these issues, this work proposes a comprehensive model for real-time production prediction based on multi-modal information fusion. Firstly, we propose to fuse image features that is extracted from indicator diagrams, with production data for the establishment of prediction models. Secondly, we develop a comprehensive model for production prediction. The model applies the long short-term memory (LSTM) network as the base model and leverages an improved artificial fish swarming algorithm (AFSA) to optimize hyper-parameters of the LSTM network. Experimental results show that (1) AFSA-LSTM model achieves high prediction accuracy, with mean absolute percentage error 4.313%; (2) our model outperforms both traditional methods and typical deep learning models; (3) predicting with multi-modal data helps our model to achieve better performances.