RESEARCH INITIATION AWARD: New Methods in Neural Network Prediction: Using Fossil Records to Improve Solar Activity Forecasting

近年来，神经网络在时间序列预测中的应用越来越受欢迎。首席研究员开发了专门的神经网络架构，该架构比传统的神经预测器显示出改进。这笔拨款将允许继续开发这些架构，同时研究下面讨论的极其重要的时间序列。最著名的时间序列之一与太阳黑子的记录相对应，可以追溯到18世纪。南澳大利亚前寒武纪时期形成的化石岩石对这个系列有了新的认识。值得注意的是，化石上的条纹宽度被认为构成了一千多年来太阳活动的记录。其形状和光谱特征与支撑系列几乎相同。这些数据虽然有明显的噪声，但信息丰富。利用神经网络，PI将研究如何利用这一系列来改进我们目前对太阳黑子的预测。一般来说，如何利用不同时间的相关序列来改进其中一个序列的预测是一个重要的问题。从实际的角度来看，太阳黑子预测的应用范围从规划轨道空间任务到保护电网。

9410823 Wan In recent years, neural networks have become increasingly popular for use in time series prediction. The principle investigator has developed specialized neural network architectures that have shown improvements over traditional neural predictors. this grant will allow continued development of these architectures while studying the extremely important times series discussed below. One of the most famous time series corresponds to the record of sunspots dating back to the 1700's. New insight in to this series come s form fossilized rocks formed during the Precambrian period in South Australia. Remarkably, striation widths in the fossils are believed to constitute over a thousand year record of solar activity. The shape and spectral characteristics are near identical to the support series. though obviously noisy, the data is rich in information. Using neural networks, the PI will research how this series may be used to improve our current predictions of sunspots. In general, it is an important problem of how related series at different times may be used to improve the prediction of one of them. From a practical perspective, the use for sunspot predictions are found in applications ranging form planning orbital space missions to safeguarding power grids.