Proving the concept of supervised machine learning to predict largest infeed and outfeed volumes and prevent over procurement of reserves – a study of Ireland’s three systems

(LSI) into Ireland’s All-island grid system. The motivation for this proof of concept (PoC) is to prevent over-procurement of system reserves by future market mechanisms. If further developed, the model’s predicted LSI values could be used in the forecasting of system reserves by serving as inputs to the system’s “Reference Incident”.

A logistic regression, implemented in Python code, was selected as the classification prediction model or “classifier”. That model was trained and tested using datasets grouped separately by calendar year (2023, 2024, and 2025). The classifier output, or “response variable”, is the predicted probability (ranging from 0 to 1) that a LSI value will fall below a specified ceiling.

Each prediction from a trained model requires a given set of input data or “explanatory variables”. These explanatory variables were found by transforming only two historical time series: demand and wind generation. Actual LSI values, necessary to train and test the model, were found from generation and interconnection time series.

Results from both the predicted classifications and model performance are very encouraging across all three calendar years. In the presented results, a 400-MW ceiling is specified upon the

All-Island system, whose maximum LSI is 500 MW. At a probability threshold of 0.900, prediction precision is extremely high at over 98%, increasing to mostly 99% at a 0.950 threshold. As expected, prediction sensitivity decreases at higher probability thresholds, but more sharply than precision increases. For example, model sensitivity drops from approximately 40% at 0.900 threshold, to approximately 10% at 0.990 threshold. Where misclassification does occur, the gap above the specified LSI ceiling was small.

The practical implication is that a trained machine-learning model can predict LSI reliably and potentially reduce the over-procurement of system reserves. The model’s performance can be optimised to balance precision and sensitivity as required to manage operational risk.

Furthermore, model development has been inexpensive, using time series data available within

EirGrid and open-source Python code.