💻 computer science

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

This paper presents a reproducible, grid-based framework and a statistically validated method for designing temporal input structures in deep learning models, which significantly improves e-scooter demand prediction accuracy in Austin, Texas, by capturing short-term persistence and cyclical patterns.

Original authors: Mohammad Sahnoon Merkebe Getachew Demissie, Roberto Souza

Published 2026-03-17

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Mohammad Sahnoon Merkebe Getachew Demissie, Roberto Souza

Original paper licensed under CC BY 4.0 (http://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Imagine you are the captain of a fleet of electric scooters in a busy city like Austin, Texas. Your job is to guess where people will want to ride next so you can move the scooters there before they arrive. If you guess wrong, you have empty scooters in quiet neighborhoods and angry people waiting in crowded areas.

This paper is like a master recipe book for making those guesses much smarter. The authors realized that while everyone was building fancy "AI brains" (deep learning models) to solve this problem, nobody had stopped to ask: "What exactly should we feed these brains, and how far back in time should they look?"

Here is the breakdown of their solution using simple analogies:

1. The Problem: The "Blurry Map" and the "Guessing Game"

Before this study, researchers were trying to predict scooter demand with two main issues:

The Blurry Map: They were taking messy trip records and trying to turn them into a picture of the city. It was like trying to paint a portrait using only a few scattered dots. They didn't have a strict, repeatable way to turn those dots into a clear, grid-based image that an AI could understand.
The Guessing Game: When deciding what past data to show the AI, they just guessed. "Let's look at the last 5 hours!" or "Let's look at the same time yesterday!" They didn't use math to prove why those specific times were the best to look at.

2. The Solution: Building a "Time-Traveling Camera"

The authors built a two-part system to fix this.

Part A: The Grid (Turning Chaos into a Picture)

Imagine the city of Austin is a giant chessboard. The authors took every single scooter trip from 2019 and dropped a "pin" on the board where the trip started and ended.

They filtered out bad data (like trips that were too short or too long).
They turned the city into a uniform grid of squares (like pixels in a photo).
They created a "heat map" for every single hour of the day, showing exactly how many people picked up or dropped off scooters in each square.
The Result: Instead of a messy list of trips, the AI now sees a clear, high-definition video of the city's scooter activity, hour by hour.

Part B: The "Time-Travel" Lens (Finding the Right Clues)

This is the paper's biggest breakthrough. They asked: "If I want to guess what happens at 5:00 PM on a Saturday, what past moments should I look at to get the best answer?"

Instead of guessing, they used a detective's magnifying glass (statistics) to find the best clues:

The Immediate Past: They looked at the hour right before (4:00 PM).
The Daily Rhythm: They looked at 5:00 PM on yesterday (because weekends often look like weekends).
The Weekly Rhythm: They looked at 5:00 PM from last Saturday (because people have habits).

They tested hundreds of combinations to see which mix of "time clues" made the AI the smartest. They found that the AI didn't just need to see the last hour; it needed to see a specific mix of recent history (what happened 1–3 hours ago) AND cyclical history (what happened at this time yesterday and last week).

3. The Experiment: The "Taste Test"

To prove their method worked, they set up a blind taste test:

Team A (The New Method): Used the scientifically chosen mix of time clues.
Team B (The "Recent" Team): Only looked at the last few hours (like checking the weather report for the last hour).
Team C (The "Fixed" Team): Only looked at the same time yesterday (like assuming today is exactly like yesterday).

The Winner: Team A crushed the competition.

For predicting the next hour, their method reduced errors by 37%.
For predicting the next 24 hours, their method reduced errors by 35%.

4. Why This Matters (The "So What?")

Think of it like cooking.

Old Way: You throw random ingredients into a pot and hope it tastes good.
New Way: You use a precise recipe that tells you exactly which spices (time clues) to use and how much of each.

The paper proves that how you prepare the data is just as important as the AI model itself. You can have the most powerful supercomputer in the world, but if you feed it the wrong "ingredients" (bad time data), it will still make bad predictions.

Summary

This paper didn't invent a new type of AI engine; instead, they built a better fuel system. They showed that by organizing city data into a clean grid and feeding the AI a scientifically proven mix of "recent" and "recurring" time patterns, we can predict where scooters are needed with much higher accuracy. This helps cities manage traffic better, reduces congestion, and makes sure you can actually find a scooter when you need one.

1. Problem Statement

The paper addresses two critical methodological gaps in deep learning-based shared micromobility (specifically e-scooter) demand prediction:

Lack of Reproducible Data Frameworks: Existing studies often treat the transformation of raw trip records into grid-based demand surfaces as a heuristic preprocessing step rather than a formally defined, reproducible pipeline. This leads to inconsistencies in spatial representation and handling of sparse data.
Arbitrary Temporal Input Design: The selection of historical time lags (input windows) for deep learning models is typically heuristic (e.g., fixed sliding windows or fixed look-back periods) rather than statistically grounded. This risks overlooking informative periodic patterns (daily/weekly cycles) or non-adjacent temporal structures, potentially limiting model generalizability.
Evaluation Flaws: Common evaluation practices often average errors over the entire spatial domain, including zero-demand areas, which dilutes performance metrics and obscures operational relevance in active zones.

2. Methodology

A. Data Processing Pipeline (Grid-Based Representation)

The authors developed a rigorous pipeline using 2019 e-scooter trip data from Austin, Texas (approx. 15 million raw records, filtered to ~4.94 million valid trips).

Filtering & Validation: Trips were filtered by duration (1–120 min), distance (0.1–35 km), speed, and valid Census Tract GEOIDs.
Spatial Discretization: Trip origins/destinations (Census Tract centroids) were projected into the UTM Zone 14N coordinate system. A uniform rectangular grid of 240m × 220m cells was constructed over the city limits.
Image Generation: Trip records were aggregated into hourly pick-up and drop-off demand images (241 × 217 pixels).
Global Activity Mask: A binary mask was generated to identify grid cells with historical activity. All training, loss calculation, and evaluation were restricted to this active region ( $\Omega$ ) to prevent error dilution from zero-demand areas.

B. Temporal Input Design Strategy

Instead of using fixed windows, the authors proposed a data-driven approach to select historical time lags:

Lag Ranking: A combined correlation-and-error-based analysis was performed on the training set.
- Correlation: Pearson correlation between target and historical demand images (same-channel and cross-channel).
- Error: Mean Absolute Error (MAE) between target and historical images.
- Ranking: Lags were ranked based on a weighted aggregation of these metrics to identify the most informative historical time steps.
Optimal Depth Selection (Ablation Study): Using a baseline UNET model, a controlled ablation study was conducted. Models were trained with varying numbers of input channels (removing the lowest-ranked lags step-by-step).
Statistical Validation:
- Non-inferiority Testing: To find the "minimal non-inferior" depth, a one-sided Wilcoxon signed-rank test was used with a 2% margin ( $\delta = 0.02 \times MSE_{min}$ ).
- Significance Testing: Pairwise comparisons between the proposed method and baselines used two-sided Wilcoxon signed-rank tests with Holm correction to control the family-wise error rate.

C. Experimental Setup

Model: Standard UNET encoder-decoder architecture (chosen as a stable baseline based on prior work).
Tasks: Next-hour prediction and Next-24-hour prediction.
Baselines:
1. Recent Adjacent History: Consecutive hours immediately preceding the target.
2. Fixed-Period History: Same hour from previous days (pure periodicity).
Metrics: Masked MSE, MAE, MaxAE, and $R^2$ (computed only over the active mask).

3. Key Contributions

Formalized Spatiotemporal Pipeline: A reproducible framework transforming raw trip data into structured hourly demand images with a global activity mask, ensuring consistent spatial learning.
Statistically Grounded Temporal Design: A methodology to identify optimal historical time lags using correlation/error analysis and determine the optimal temporal depth via non-parametric hypothesis testing, moving beyond heuristic window selection.
Rigorous Evaluation Protocol: Implementation of paired non-parametric statistical testing (Wilcoxon signed-rank with Holm correction) to validate that performance improvements are statistically significant, not just numerical.

4. Results

Optimal Temporal Configurations

Next-Hour Prediction: The optimal configuration required 36 input channels (18 pick-up + 18 drop-off lags). The selected lags included immediate history (1–3 hours) and periodic anchors (23–25, 168, 336, 504 hours). Reducing depth below this caused statistically significant performance degradation.
Next-24-Hour Prediction: A 18-channel configuration (9 pick-up + 9 drop-off lags) was found to be minimal non-inferior to the best-performing (26-channel) model. The lags focused heavily on daily and weekly periodicity (24, 48, 168, 336 hours).

Comparative Performance (Test Set)

The proposed method significantly outperformed both baselines:

Next-Hour:
- MSE Reduction: Reduced MSE by ~26% compared to the "Recent Adjacent" baseline (8.45 vs. 11.40) and ~37% compared to the "Fixed-Period" baseline.
- Statistical Significance: Holm-corrected p-values were extremely low ( $< 10^{-40}$ ), confirming the superiority of the proposed lag selection.
Next-24-Hour:
- MSE Reduction: Reduced MSE by ~35% compared to the "Recent Adjacent" baseline (26.37 vs. 40.46) and ~14% compared to the "Fixed-Period" baseline.
- Insight: For the 24-hour horizon, the "Fixed-Period" baseline outperformed the "Recent Adjacent" baseline, indicating that long-horizon prediction relies more on daily/weekly cycles than immediate momentum.

5. Significance

Principled Input Design: The study demonstrates that carefully structured temporal inputs can yield substantial performance gains (up to 37% MSE reduction) even without changing the underlying deep learning architecture.
Operational Relevance: By using an activity mask, the results reflect performance in actual usage zones, providing more actionable insights for fleet rebalancing and management than global error metrics.
Methodological Rigor: The paper sets a new standard for evaluating spatiotemporal models by mandating statistical validation of performance differences, addressing the "black box" nature of heuristic tuning in previous literature.
Generalizability: The framework is designed to be reproducible and applicable to other grid-based urban micromobility systems, provided similar data structures exist.

Conclusion: The paper concludes that while deep learning architectures are important, the design of the temporal input structure is a foundational determinant of predictive success. Future work should focus on integrating these optimized inputs with regime-aware architectures (e.g., accounting for weather or events) to further enhance long-horizon prediction.