Features/failure forecast

[Oklyagina]

Failure forecasting integration

Failure forecasting is integrated directly into the existing PowerGrid simulator pipeline.
Execution flow

1. Forecasting artifacts are loaded when the simulator starts: app/models/Simulator.py:102-117.
2. Grid observations are recorded every step to maintain the required history: app/models/Simulator.py:197-200.
3. When an overload is detected, the original overload event is sent: app/models/Simulator.py:354-365.
4. After InteractiveAI returns the overload event ID, failure forecasting is executed for the overloaded line: app/models/Simulator.py:366-378.
5. A linked forecast event is sent using the overload ID as its parent_event_id: app/models/Simulator.py:379-387.
   Grid observation
   → Overload detected
   → Overload event sent
   → Failure forecast executed
   → Linked forecast event sent
   The inference pipeline and forecast output are implemented in app/models/failure_forecast.py:238-301.

The output includes:

• binary failure prediction (0 - no fail, 1 - fail);
• failure probability;
• forecast horizon;
• overloaded line ID.

The linked event payload is created in app/models/failure_forecast.py:304-346 and published through the simulator’s existing event mechanism in app/models/Communicate.py:515-533.

Getting pre-trained model:

The train models are available as artifacts in the latest release here: https://github.com/ainetus/RL_agent_failure_forecast/releases/tag/v2.0-docker

It is required to download ai4realnet_small-curriculum-v2026.09.18-models.zip archive and unzip the failure_forecaster subdirectory to the project root.

[Oklyagina]

My commits are only related to the forecasting. Other ones were there before, but were not merged to the main