Large geomagnetically induced currents (GICs) pose a risk to ground based infrastructure such as power networks. Large GICs may be induced when the rate of change of the ground magnetic field is significantly elevated. We assess the ability of three different machine learning model architectures to process the time history of the incoming solar wind and provide a probabilistic forecast as to whether the rate of change of the ground magnetic field will exceed specific high thresholds at a location in the UK. The three models tested represent feed forward, convolutional and recurrent neural networks. We find all three models are reliable and skillful, with Brier skill scores, receiver-operating characteristic scores and precision-recall scores of approximately 0.25, 0.95 and 0.45, respectively. When evaluated during two example magnetospheric storms we find that all scores increase significantly, indicating that the models work better during active intervals. The models perform excellently through the majority of the storms, however they do not fully capture the ground response around the initial sudden commencements. We attribute this to the use of propagated solar wind data not allowing the models notice to forecast impulsive phenomenon. Increasing the volume of solar wind data provided to the models does not produce appreciable increases in model performance, possibly due to the fixed model structures and limited training data. However, increasing the horizon of the forecast from 30 min to 3 h increases the performance of the models, presumably as the models need not be as precise about timing.

中文翻译：

---

预测英国地磁场大变化率的概率：时间尺度、视野和阈值

大地磁感应电流 (GIC) 对电力网络等地面基础设施构成风险。当地面磁场的变化率显着升高时，可能会诱发大 GIC。我们评估了三种不同的机器学习模型架构处理入射太阳风的时间历史的能力，并提供关于地面磁场变化率是否会超过英国某个地点的特定高阈值的概率预测。测试的三个模型分别代表前馈、卷积和循环神经网络。我们发现所有三个模型都是可靠和熟练的，Brier 技能分数、接收器操作特征分数和精确召回分数分别约为 0.25、0.95 和 0.45。在两个示例磁层风暴中进行评估时，我们发现所有分数都显着增加，表明模型在活动间隔期间工作得更好。这些模型在大多数风暴中表现出色，但它们并没有完全捕捉到最初突然开始周围的地面响应。我们将此归因于传播的太阳风数据的使用，不允许模型注意到预测脉冲现象。增加提供给模型的太阳风数据量不会显着提高模型性能，这可能是由于固定的模型结构和有限的训练数据。然而，将预测范围从 30 分钟增加到 3 小时会提高模型的性能，这大概是因为模型在时间方面不需要那么精确。