如何预测时间序列中的季节性峰值，而忽略非季节性的峰值？

Question

我刚刚开始进行时间序列预测，并试图为后续用例找出解决方案。 我想检测到非季节性警报进入我们的系统。如果传入警报是季节性的，我想忽略它们。不适合季节性模式的异常值，我需要将其升级到处理模块。

#Creating time series which has spikes every 20th time interval.
alert_once_a_day = [1.0 if i % 20 == 0 else 0.0  for i in range(100)]
#Adding an outlier at 27, which does not fit pattern of spikes at every 20 th interval.
alert_once_a_day[27] =1.0

在以上系列中，我想查找所有警报以季节性模式出现并忽略它们。

Answer 1

最简单的答案，请使用遮罩：

alert_once_a_day = [1.0 if i % 20 == 0 else 0.0  for i in range(100)]

seasonal_alerts = list(alert_once_a_day)
mask = np.array(seasonal_alerts) == 0

#Adding an outlier at 27, which does not fit pattern of spikes at every 20 th interval.
alert_once_a_day[27] =1.0

# assuming your zeros are "non alert" values, the mask will eliminate seasonal peaks:    
true_alerts = np.where(np.array(alert_once_a_day) * mask==1)
true_alerts
>> (array([27], dtype=int64),)