I have a multivariate time series anomaly detection problem where data is time series and some of the variables follow specific patterns depending on time.
For example, let's say variables A, B and C take high but consistent values every 60 minutes, which would look like anomaly from the perspective of other time windows, however knowing the context of the data, I know that this is not anomalous, because these very high values occur periodically and are expected. However it would be anomaly if it were to occur in other timeframe rather than every 60 minutes.
Therefore I want to include this knowledge of 'time' and 'periodicity' in my model, so that these periodic instances will not be marked as anomaly when they occur in periodicity but will be marked as anomaly other times.
Since I have 2k features, I don't think simply adding date/time as another column can help. So what is the best way to model this kind of time dependent anomaly detection where periodicity exists?