Risk-Aware Load Forecasting for DC Trolleybus Networks Using an LSTM with Monte Carlo Dropout

Modern trolleybus networks are increasingly integrated with energy management strategies, energy storage systems, and smart-grid functionalities. Accurate short-term forecasting of substation demand is essential for operational planning, yet traditional approaches rely on detailed physical models of the traction network, vehicle dynamics, and traffic schedules. This paper presents a lightweight and fully datadriven method based on a multi-horizon Long Short-Term Memory (LSTM) architecture with Monte Carlo dropout for probabilistic short-term forecasting in DC trolleybus substations. The model uses only historical voltage and current measurements recorded at the substation, enriched with temporal features, without requiring timetable information, vehicle-level models, or network topology. Real-world data from multiple substations of the Bologna (Italy) trolleybus system are used to train unified encoder-decoder LSTM models for prediction horizons between 5 and 120 minutes. Experimental results show accurate point forecasts and useful uncertainty intervals, with narrow prediction bands suitable for risk-aware operational planning. The proposed approach complements physics-based simulators by providing a scalable, low-overhead tool for real-time load anticipation, energy management, and potential integration with storage and demandresponse strategies in DC transit networks.