Community-Structured CNN-LSTM with Dynamic Weather Attention for Bike-Sharing Demand Forecasting Across Multiple Cities

Accurate short-horizon demand forecasting is essential for efficient bike-sharing rebalancing operations, yet most existing approaches validate on a single city and predict pickup demand only, leaving open questions about generalizability and the joint modeling of departure, arrival, and net supply flows. This paper proposes a CNN-LSTM architecture augmented with two components: a dynamic weather attention gate that conditions model output on forecast-horizon weather conditions and a community graph integration module that diffuses spatial context through a trip-volume-weighted adjacency matrix derived from Leiden community detection. A four-variant ablation study isolates the contribution of each component across two geographically and climatically distinct bike-sharing systems, BIXI Montreal and Capital Bikeshare Washington DC, using 15 min resolution trip data from two consecutive riding seasons. A single-seed evaluation initially suggested that community graph integration consistently reduces pickup prediction error in both cities and that the weather attention gate improves pickup prediction in Washington DC but not Montreal. However, a subsequent multi-seed check (three random seeds) found that these improvements do not hold up as consistently as the single-seed result suggested: The community graph variant outperforms the base CNN-LSTM in only 25–47% of seed-community combinations across the two cities and the weather attention variant in only 27–60%, indicating that small percentage improvements reported from a single training run in this class of model are frequently within the range of ordinary seed-to-seed variations rather than reliable architectural effects. We report this directly as a methodological finding in its own right: Ablation studies at the scale typically reported in this literature, including our own initial single-seed results, may substantially overstate the reliability of small reported improvements. For net supply change, the quantity most directly relevant to rebalancing decisions, all deep learning variants significantly outperform Random Forest in Montreal (paired Wilcoxon p=0.0005), a pattern that trends similarly but does not reach significance in Washington DC at the available sample size (p≈0.06–0.07); no significant difference is observed between the base CNN-LSTM and the full proposed model in either city, indicating that the added architectural components do not provide a demonstrable further benefit specifically for this derived, signed target. Taken together with the multi-seed instability observed for pickup prediction, these results caution against over-interpreting small single-seed ablation margins in this modeling setting more broadly and point to multi-seed evaluation as necessary practice for this class of architecture comparison.