AI Use on Railway Energy Consumption

The Belgian railway infrastructure manager, Infrabel, foresees the necessary infrastructure for the trains to drive safely on the Belgian railway network. This includes the distribution of energy towards the trains, but also to the fixed assets necessary for said trains to operate (datacentres, track changes, heating, stations, …). The Belgian railway is among the largest purchaser of electricity in Belgium with an average yearly consumption close to 1.5TWh. The energy distribution on a weekly basis fluctuates based on the rush hours (that are very distinguishable), where smaller peaks are noticeable during the weekends. Despite the clear pattern, it still remains difficult to make a traditional forecast model since for a given quarter hour the consumption may vary up to 30%, with even larger variances during exceptional events.

Due to energy transition and the penetration of renewables, accurate predictions of our consumption becomes primordial in order to avoid excessive costs of balancing the Belgian electricity grid. Imbalance costs are difficult to predict and can work both ways (positive/negative) for purchasing additional energy as well as selling excess amounts. Given the big volumes the railway is exposed to, these imbalance costs can quickly gain large proportions. As such, the careful approach is to predict the consumption as accurately as possible whilst mitigating the imbalance risks of the energy market. The preference is given to handle this job via a model built with Artificial Intelligence. Thanks to supervised machine learning, a promising model was ascertained that can predict the energy consumption with an average error under 6%. Traditional models would insufficiently capture the complexity of the market and hence not shield us adequately against fluctuations.