Neural networks to analyze uncertainty in prognostic model runs

Introduction

Current prediction methods strive to minimize the difference or the error between the prediction result and the actual value. Extrapolating historical data records into the future always results in increasing error, and the prediction can only be applied into a short time outreach. This is because data contain uncertainty, and no current methods take into account the uncertainty wedge which always increases with time. Projection is therefore needed to capture possible values in the future. Furthermore, we wish to maintain a balance between the historical dynamics, the opening, and the extension of uncertainty wedge in the prognostic outreach.

Methods

We define a method of constructing the uncertainty wedge. The historical data is divided into three parts as following: i) the L1 part to capture historical dynamics; ii) the L2 part to capture dynamics change and construct the uncertainty wedge; iii) the testing part to apply the wedge with the historical data. We use Radial Basis Function (RBF) Neural Networks which has reported good performance in short-term prediction. Changing the capture of dynamics affects a different uncertainty wedge and length of extension. We then manipulate the complexity of capturing the dynamics by changing the number of nodes in the network to see the effect on the uncertainty wedge and the extension.

Results and conclusions

We apply the method of constructing an uncertainty wedge for climate-related data (time series and non-time series), and investigate how complexity affects the prognostic outreach of the model. To conclude, models that are too simple or too complex both have poor projective performance, and it is possible to optimize complexity to reach the optimal projection.

Please note these Proceedings have received limited or no review from supervisors and IIASA program directors, and the views and results expressed therein do not necessarily represent IIASA, its National Member Organizations, or other organizations supporting the work.

Last edited: 03 February 2016