Description

The ADI Prediction server is a standard ADI server offering a service allowing:

• A client application to load models that are visible to any or selected applications.
• Many applications to request concurrently different predictions from the same or multiple models.
• Many models to reside at the same time in the ADI Prediction server cache, speeding up the prediction evaluation.
• The pre-loaded models definitions to be discovered by navigating the Address Space of the ADI Prediction server.

Main success scenario

Scenario steps	Interface used
The client application logs on ADI Prediction server	OPC UA CreateSession
The client application reads model from disk and loads them in ADI Prediction server	OPC UA AddNode and write services
ADI Prediction server exposes all model structures in its address space	OPC UA address space + ADI data types
The client application logs off ADI Predictor server	OPC UA CloseSession
The same or another client application logs on ADI Prediction server	OPC UA CreateSession
Client application identifies loaded models and their structure	OPC UA Browse address space + ADI data types
Client application calls MVAPredict () method with the appropriate input parameters,	OPC UA/ADI method call service + ADI data types
ADI Prediction server computes predicted values using vendor native libraries.	Vendor specific native libraries
Application reads predicted values from MVAPredict () method output parameters.	OPC UA/ADI method call service + ADI data types
Repeat steps 7 to 9 for each model
Application logs off from ADI Prediction server	OPC UA CloseSession

Many applications may execute Steps 5 to 11 concurrently.

1. The client application may pass the input values and received predicted values using OPC UA/ADI address space. However, this approach does not scale up well in multi-user environment because a prediction server is expected to serve many users concurrently which make synchronization almost impossible.
2. In custom / dedicated configurations, the ADI Prediction server loads a predefined set of models, so client applications can use them immediately.