Toward an Adaptive Artificial Neural Network–Based Postprocessor

• Published in: Monthly weather review Vol. 149; no. 12; pp. 4045 - 4055
• Main Author: Roebber, Paul J
• Format: Journal Article
• Language: English
• Published: Washington American Meteorological Society 01.12.2021
• Subjects: Algorithms; Artificial neural networks; Back propagation; Back propagation networks; Computer applications; Computer architecture; Data storage; Dynamic models; Forecasting data; Information storage; Modelling; Neural networks; Numerical prediction; Numerical weather forecasting; Predators; Prediction models; Prey; Regression analysis; Storage requirements; Training; Weather forecasting
• DOI: 10.1175/MWR-D-21-0089.1
• ISSN: 0027-0644

Abstract We introduce an adaptive form of postprocessor where algorithm structures are neural networks where the number of hidden nodes and the network training features evolve. Key potential advantages of this system are the flexible, nonlinear mapping capabilities of neural networks and, through backpropagation, the ability to rapidly establish capable predictors in an algorithm population. The system can be implemented after one initial training process and future changes to postprocessor inputs (new observations, new inputs, or model upgrades) are incorporated as they become available. As in prior work, the implementation in the form of a predator–prey ecosystem allows for the ready construction of ensembles. Computational requirements are minimal, and the use of a moving data window means that data storage requirements are constrained. The system adds predictive skill to a demonstration dynamical model representing the hemispheric circulation, with skill competitive with or exceeding that obtainable from multiple linear regression and standard artificial neural networks constructed under typical operational limitations. The system incorporates new information rapidly and the dependence of the approach on the training data size is similar to multiple linear regression. A loss of performance occurs relative to a fixed neural network architecture in which only the weights are adjusted after training, but this loss is compensated for by gains from the ensemble predictions. While the demonstration dynamical model is complex, current numerical weather prediction models are considerably more so, and thus a future step will be to apply this technique to operational weather forecast data.