Inter-Stop Energy Prediction and Causal Driver Quantification for Dual-Source Trolleybuses via a Time-Aware Tabular Deep Learning Architecture

Dual-source trolleybuses (DTBs) offer a flexible alternative to battery-electric buses by switching between overhead catenary power and on-board batteries. However, their energy consumption is highly complex, influenced by a mix of static route attributes, high-frequency vehicle trajectories, and fluctuating meteorological conditions. This paper addresses the challenge of accurately predicting inter-stop energy use and, crucially, moving beyond black-box predictions to identify the causal factors that drive energy consumption.

The authors propose a two-part framework. First, they introduce a "Time-Aware TabM" architecture that embeds periodic time encoding (Time2Vec) into a parameter-efficient batch-ensemble backbone. This allows the model to jointly process heterogeneous inputs—static tabular data and dynamic temporal sequences—within a single network. The model is optimized using Bayesian optimization with tree-structured density estimation to ensure efficiency. Second, the authors implement a three-layer causal pipeline: SHAP values quantify marginal feature contributions, DirectLiNGAM recovers the causal structure (directed acyclic graph), and a T-Learner meta-algorithm estimates the net average treatment effects of specific operational interventions.

The proposed model achieved a Mean Absolute Percentage Error (MAPE) of 6.52% and an R-squared of 0.982 on the Zurich trolleybus dataset, outperforming ten baseline models including traditional statistical methods and standard deep learning architectures. Ablation studies confirmed that the periodic time encoding was the most significant contributor to predictive accuracy. The causal analysis revealed that the regenerative braking ratio and average speed are the most effective levers for energy saving, while coasting distance is the primary driver of excess energy consumption. These insights provide actionable thresholds for optimizing driving behavior, vehicle technology, and catenary network planning.