Time series forecasting with missing data using neural network and meta-transfer learning

In this paper, meta-learning and transfer-learning are combined to solve the missing data problem in time series forecasting. We successfully solved the common problems of neural networks based time series forecasting: local minimum and the poor predication accuracy with missing data. The strong and weak convergence of the proposed algorithms are analyzed. We successfully applied the proposed method to predict the air pollution of Mexico City with missing data. The prediction results and comparison show that the meta-transfer learning has very good prediction accuracy, when the air monitoring network has some failures.