Univ. of Alaska-Fairbanks: Hybrid neural network/expert systems for mesoscale weather forecasting: Transferring technology to the operational forecaster

Univ. of Alaska-Fairbanks: "Hybrid neural network/expert systems for mesoscale weather forecasting: Transferring technology to the operational forecaster"

Final Report

This Partners Project with the NWS Fairbanks office was a follow-up to a previous project on using neural networks, along with expert systems, to predict lightning in certain regions of Alaska during the summer months. This later project had four main objectives:

To investigate Counterpropagation Networks (CPN) as an alternative to the Backpropagation Networks (BPN) that had previously been used for lightning prediction, as they offer faster training time and it is possible to correlate the network's prediction in a particular case with the classification exemplar the network used for that prediction
To reconsider the earlier work done on a hybrid expert system for lightning prediction and to integrate the information gained from the CPN's exemplar for a prediction with knowledge from an expert system to improve overall system performance. A further goal on this part of the project was to determine how to integrate the point-based expert system knowledge with the area-based neural network prediction
To integrate the research work into operational forecasting
To conduct a workshop to teach weather service personnel some of the science behind artificial intelligence technologies and to provide a forum for discussing practical problems associated with the use of these technologies

In relation to the first objective, a CPN was trained to predict lightning. The performance of the best CPN was better than that of the best BPN, and the researchers concluded that CPN is an effective network technology. Work on the second objective was not complete when the Partners Project ended but is continuing. As part of the third objective, the researchers used data for training the CPN which was not preselected or prescreened in contrast to what was done on the previous study using the BPN. The CPN's performance statistics on this data showed the robustness of the network in an operational environment. Additionally, the system output was modified and integrated with ARONET, the Alaska Region Operational NETwork, which is the main interface tool to forecast data used by Fairbanks forecasters. Finally, the workshop was held during the week of June 12-16, 1995.