Thanks for your question and interest in this topic! For a deeper insight into how the algorithm works, I would like to refer you to DataMiner Prediction engine - DataMiner Dojo I believe this will contain a lot of information that you will find useful.
Now, let me answer more specifically to your question.
Given a parameter, we typically don’t make just one prediction, but 4 different predictions. The first one is based on about 2500 points of real-time data: this prediction is the most accurate on the short term, but can not be used to predict very far in time. The second/third is based on about 2500 points of 5’/1h average data fetched from the database and better captures the long term behavior of the parameter. The fourth prediction finally is based on daily average data, which is obtained by averaging the 1h averages from the database. This one really takes these long term behaviors into account.
Apart from trend data, our prediction engine also uses the change points (anomalies) that DataMiner discovered in your parameter. (see Working with behavioral anomaly detection | DataMiner Docs) This is used in the preprocessing step, where the trend data is first “cleaned” before we build the actual prediction model. For example, when an irregular shift in level occurs in the data, then this distorts the data in such a way that it is typically very difficult to create an accurate model on the entire data set. So, we account for this by basically “removing” this level shift during the modelling step. Similarly, we take actions to deal with unexpected outliers, trend changes,…
Of course, we don’t take any drastic actions that ignore periodicity in the data. During the preprocessing state, our algorithm looks for repeating behaviors. Typically, if it occurs at least four times in the trend data, then it is taken into account.
The forecasting algorithm has grown so much over time and has become so sophisticated that it’s impossible to explain everything fully in this post. If you have any additional questions or you want to discuss this further, don’t hesitate to post a follow up or request a call in this thread and we can certainly discuss further on teams.