cs.CVMar 13, 2026

A Grid-Based Framework for E-Scooter Demand Representation and Temporal Input Design for Deep Learning: Evidence from Austin, Texas

Authors: Mohammad Sahnoon, Merkebe Getachew Demissie, Roberto Souza

Organizations: Department of Electrical and Software Engineering, Schulich School of Engineering, University of Calgary, Calgary, AB T2N 1N4, Canada · Department of Civil Engineering, Schulich School of Engineering, University of Calgary, Calgary, AB T2N 1N4, Canada

Read PDF
Ask AI
Cite

Copy citation

DOINot available for this paper

Abstract

Despite progress in deep learning for shared micromobility demand prediction, the systematic design and statistical validation of temporal input structures remain underexplored. Temporal features are often selected heuristically, even though historical demand strongly affects model performance and generalizability. This paper introduces a reproducible data-processing pipeline and a statistically grounded method for designing temporal input structures for image-to-image demand prediction. Using large-scale e-scooter data from Austin, Texas, we build a grid-based spatiotemporal dataset by converting trip records into hourly pickup and dropoff demand images. The pipeline includes trip filtering, mapping Census Tracts to spatial locations, grid construction, demand aggregation, and creation of a global activity mask that limits evaluation to historically active areas. This representation supports consistent spatial learning while preserving demand patterns. We then introduce a combined correlation- and error-based procedure to identify informative historical inputs. Optimal temporal depth is selected through an ablation study using a baseline UNET model with paired non-parametric tests and Holm correction. The resulting temporal structures capture short-term persistence as well as daily and weekly cycles. Compared with adjacent-hour and fixed-period baselines, the proposed design reduces mean squared error by up to 37 percent for next-hour prediction and 35 percent for next-24-hour prediction. These results highlight the value of principled dataset construction and statistically validated temporal input design for spatiotemporal micromobility demand prediction.

Figures & tables

Fig. 1: Trip counts per vehicle type in Austin’s raw micromobility dataset (15 million samples).
Filter CategoryCondition Applied
Vehicle typeRetain e-scooter trips only.
Trip yearRetain trips occurring within the 2019 calendar year.
Trip durationKeep trips between 1 and 120 minutes inclusive.
Trip distanceKeep trips between 0.1 and 35 kilometers inclusive.
Average speedKeep trips with speed between 2 and 26 kilometers per hour inclusive.
Tract GEOIDTrip origin and destination must have valid census tract identifiers.
TABLE I: Filtering criteria applied to construct the final Austin, TX 2019 e-scooter dataset.
Fig. 2: Spatial coverage of the final 2019 e-scooter dataset. The figure displays all 205 unique census tracts (in black) that appear as trip origins or destinations, along with the Austin city limits (in red) used for spatial filtering.
Fig. 3: Distribution of core trip characteristics for the processed 2019 e-scooter dataset: (a) trip duration in minutes, (b) trip distance in kilometers, and (c) trip average speed in kilometers per hour.
Fig. 4: Average daily e-scooter demand in 2019, obtained by averaging daily trip counts across all days of the same weekday.
Fig. 5: Average hourly e-scooter demand in 2019, calculated by averaging trip counts for each hour of the day across all calendar days.
Fig. 6: Spatial distribution of (a) trip origin and (b) trip destination intensities for Austin’s 2019 e-scooter dataset at the census tract level, highlighting high-demand tracts in central Austin, TX.
Fig. 7: Generation of the 240 m × 220 m240\,m\,\times\,220\,m rectangular grid over the selected census tracts within the Austin, TX municipal boundary.
Fig. 8: Sample pick-up demand image generated for 4:00 p.m. on March 17, 2019. Subplot (a) uses pixel-based indexing of the raster grid, subplot (b) expresses the same demand field in projected UTM Zone 14N coordinates, and subplot (c) shows the demand field in geographic coordinates. The color legend indicates the ground-truth demand level associated with each pixel value and is included for visualization only. For model input, demand images are represented in black and white, where the intensity of each white pixel encodes the demand level.
Fig. 9: Global binary activity mask used to define the operational footprint of the historically active e-scooter demand locations to guide model training and evaluation across both the pick-up and drop-off channels. Subplot (a) uses pixel-based indexing of the raster grid, subplot (b) expresses the same demand field in projected UTM Zone 14N coordinates, and subplot (c) shows the demand field in geographic coordinates.
Fig. 10: Channel-based input representation for spatiotemporal demand prediction. The figure illustrates the pick-up demand case, where historical demand images are stacked as ordered channels (e.g., Ch. 1,Ch. 2,…,Ch. n\mathrm{Ch.}~1,\mathrm{Ch.}~2,\ldots,\mathrm{Ch.}~n ). The same construction is extended to drop-off demand by appending the corresponding channels and producing a two-channel output, which is evaluated against the ground-truth target map. Pixel color coding is for demand-level visual illustration only.
Next-hourNext 24-hour
Rank final ( τ\tau )τ\tau (hours)Rank final ( τ\tau )τ\tau (hours)
11124
42325
1731026
1122948
52312143
TABLE II: Top 20 selected historical time lags (in hours) for next-hour and next 24-hour e-scooter demand prediction. Each lag τ\tau corresponds to demand observed at time (t−τ)(t-\tau) .
Prediction HorizonConfiguration TypeChannelsTime-lag set τ\tau (hours)
Next-hourProposed method36{1,2,3,22,23,24,25,26,48,143,144,145,167,168,169,192,336,504}\{1,2,3,22,23,24,25,26,48,143,144,145,167,168,169,192,336,504\}
Recent adjacent history36{1,2,3,…,17,18}\{1,2,3,\ldots,17,18\}
Fixed-period history36{1,25,49,73,97,121,145,169,193,217,241,265,289,313,337,361,\{1,25,49,73,97,121,145,169,193,217,241,265,289,313,337,361,
385,409}385,409\}
Next 24-hourProposed method18{24,25,48,144,167,168,169,192,336}\{24,25,48,144,167,168,169,192,336\}
Recent adjacent history18{24,25,26,27,28,29,30,31,32}\{24,25,26,27,28,29,30,31,32\}
TABLE III: Historical Time-lag configurations used to validate the proposed historical time-lag selection method under the next-hour and next 24-hour prediction horizons. For each configuration, the same set of time lags τ\tau is applied identically to pick-up and drop-off demand, resulting in separate input channels with a shared temporal structure.
Prediction HorizonConfigurationInput ChannelsMSE*Time Lags ( τ\tau ) (See Table II )
Next-hourMinimum MSE3620.11Top 18
Minimal non-inferiorSame as minimum-MSE configuration
Next 24-hourMinimum MSE2661.20Top 13
Minimal non-inferior1863.42Top 9
TABLE IV: Optimal historical input configurations identified for each prediction horizon using the UNET model. The table reports the validation-set minimum-MSE configuration and the minimal non-inferior configuration determined through Holm-corrected non-inferiority testing.
ConfigurationMSEMAER2R^{2}MaxAE
Proposed method8.450.500.7824.33
Recent adjacent history11.400.570.5734.51
Fixed-period history13.480.610.5538.03
TABLE V: Overall test-set performance of temporal input (time-lag) selection methodologies for next-hour e-scooter demand prediction using the UNET model, computed within the masked active region and reported as metrics averaged jointly over pick-up and drop-off demand channels.
ConfigurationMSEMAER2R^{2}MaxAE
Proposed method26.370.75-0.7750.47
Recent adjacent history40.460.87-1.7164.46
Fixed-period history30.860.85-0.3653.26
TABLE VI: Overall test-set performance of temporal input (time-lag) selection methodologies for next 24-hour e-scooter demand prediction using the UNET model, computed within the masked active region and reported as metrics averaged jointly over pick-up and drop-off demand channels.
HorizonComparisonBetter ConfigurationHolm corrected p-value
Next-hourProposed method vs. recent adjacent historyProposed method5.78×10−405.78\times 10^{-40}
Proposed method vs. fixed-period historyProposed method1.43×10−751.43\times 10^{-75}
Recent adjacent vs. fixed-period historyRecent adjacent history1.89×10−41.89\times 10^{-4}
Next 24-hourProposed method vs. recent adjacent historyProposed method1.15×10−351.15\times 10^{-35}
Proposed method vs. fixed-period historyProposed method1.55×10−81.55\times 10^{-8}
Recent adjacent vs. fixed-period historyFixed-period history2.53×10−322.53\times 10^{-32}
TABLE VII: Holm-corrected pairwise statistical comparison between historical time-lag selection methodologies on test MSE for next-hour and next 24-hour e-scooter demand prediction. Results are based on two-sided Wilcoxon signed-rank tests applied to per-sample MSE distributions averaged jointly over pick-up and drop-off demand.
Figure 18
Figure 19
Figure 20
More in this topic

Spatiotemporal Forecasting

Spatiotemporal forecasting predicts how quantities change across locations and time, from traffic and weather to disease and environmental systems.

See all 423 papers

Explore similar work

CardsList
Feb 18, 2026

MoDE-Boost: Boosting Shared Mobility Demand with Edge-Ready Prediction Models

Urban demand forecasting plays a critical role in optimizing routing, dispatching, and congestion management within Intelligent Transportation Systems. By leveraging data fusion and analytics techniques, traffic demand forecasting serves as a key intermediate measure for identifying emerging spatial and temporal demand patterns. In this paper, we tackle this challenge by proposing two gradient boosting model variations, one for classiffication and one for regression, both capable of generating demand forecasts at various temporal horizons, from 5 minutes up to one hour. Our overall approach effectively integrates temporal and contextual features, enabling accurate predictions that are essential for improving the efficiency of shared (micro-) mobility services. To evaluate its effectiveness, we utilize open shared mobility data derived from e-scooter and e-bike networks in five metropolitan areas. These real-world datasets allow us to compare our approach with state-of-the-art methods as well as a Generative AI-based model, demonstrating its effectiveness in capturing the complexities of modern urban mobility. Ultimately, our methodology offers novel insights on urban micro-mobility management, helping to tackle the challenges arising from rapid urbanization and thus, contributing to more sustainable, efficient, and livable cities.

Antonios Tziorvas, George S. Theodoropoulos, Yannis Theodoridis

Department of Informatics, University of Piraeus, Piraeus, Greece.

Spatiotemporal ForecastingIntelligent Transportation SystemsGradient-Boosted Decision Trees
AskPDF
Jul 7, 2026

STAGformer: A Spatio-temporal Agent Graph Transformer for Micro Mobility Demand Forecasting

Accurate station-level demand forecasting is essential for the efficient operation of bike-sharing systems, yet it remains challenging due to complex spatio-temporal dependencies and the large scale of urban networks. This paper presents STAGformer, a Spatio-Temporal Agent Graph Transformer that achieves efficient global modeling with linear computational complexity. The model introduces a two-step agent attention mechanism, where a small set of learnable spatial and temporal agent tokens first aggregate global information and then broadcast it back to individual stations and time steps, effectively capturing long-range interactions while reducing the quadratic cost of standard self-attention to O(NT). STAGformer integrates four core modules: a spatio-temporal encoder that fuses dynamic node features with external contextual factors (weather, time, points of interest), a graph propagation module for spatial neighbor aggregation, a temporal convolution module for local pattern extraction, and the agent attention module for global dependency modeling. Extensive experiments on two real-world datasets -- NYC Citi-Bike and Chicago Divvy-Bike -- demonstrate that STAGformer consistently outperforms state-of-the-art baselines across multiple prediction horizons, achieving significant improvements in both RMSE and MAE. Ablation studies validate the contribution of each component, with the agent attention mechanism proving critical for modeling global spatio-temporal dependencies.

Ye Zihao

Department of System Engineering City University of Hong Kong

Attention MechanismsSpatiotemporal ForecastingTime Series TransformersSpatiotemporal GNNs
AskPDF
Oct 2, 2026

LLM-enhanced spatio-temporal learning for grid-level docked bike sharing demand prediction

Short-term bike-sharing demand forecasting is complicated by spatial-temporal non-stationarity and the practical difficulty of incorporating unstructured external text into numerical pipelines. Conventional approaches rely on historical flow sequences and fixed graph structures, thereby constraining their accuracy when anomalous social events perturb normal travel patterns. We propose a forecasting framework in which a Large Language Model (LLM) drives a semantic shockwave mechanism that converts free-form urban text, such as municipal event schedules, local news, and transit bulletins, into quantified spatial-temporal perturbation fields. The LLM extracts three physically interpretable parameters per event (intensity, spatial reach, and temporal lag), from which Gaussian decay fields are constructed and injected into a Zero-Inflated Adaptive Spatio-Temporal Graph Convolutional Network (ZI-ASTGCN). To handle the pronounced sparsity of grid-level measurements, the model couples a dual-branch output head with a multi-task zero-inflated loss that jointly trains a gating probability and a conditional flow intensity. Experiments on the operational Barcelona Bicing dataset show that ZI-ASTGCN outperforms established neural baselines, with particularly strong gains during high-demand periods, validating the utility of physics-grounded semantic signals in spatial-temporal mobility forecasting.

Xuxilu Zhang, Francesc Soriguera

BIT – Universitat Polit`ecnica de Catalunya – Barcelona Tech (UPC), Jordi Girona 1–3, Building B1, Office 111, Barcelona 08034, Spain.

Spatiotemporal ForecastingTime Series ForecastingSpatiotemporal GNNs
AskPDF