Visualization of Neural Network Predictions for Weather Forecasting

Abstract

Recurrent neural networks are prime candidates for learning relationships and evolutions in multi-dimensional time series data. The performance of such a network is judged by the loss function, which is aggregated into a single scalar value that decreases during successful training. Observing only this number hides the variation that occurs within the typically large training and testing data sets. Understanding these variations is of highest importance to adjust hyperparameters of the network, such as the number of neurons, number of layers or even to adjust the training set to include more representative examples. In this paper, we design a comprehensive and interactive system that allows to study the output of recurrent neural networks on both the complete training data as well as the testing data. We follow a coarse-to-fine strategy, providing overviews of annual, monthly and daily patterns in the time series and directly support a comparison of different hyperparameter settings. We applied our method to a recurrent convolutional neural network that was trained and tested on 25 years of climate data to forecast meteorological attributes, such as temperature, pressure and wind speed. The presented visualization system helped us to quickly assess, adjust and improve the network design.