Intelligent Scheduling and Pressure‑Balancing System at Ele.me: Machine‑Learning Applications

Instant delivery logistics is the core value of the food‑delivery industry, where orders must be fulfilled within about 30 minutes, posing significant challenges for algorithmic models. This article presents Ele.me's intelligent scheduling system, emphasizing the pressure‑balancing subsystem and illustrating the role of machine‑learning algorithms in solving real‑world O2O problems.

1. Ele.me Intelligent Scheduling System

The system automates most dispatch tasks, reducing human intervention and achieving automated, intelligent order assignment. It consists of four sub‑systems (see Figure 1):

Intelligent Dispatch: matching riders with orders and planning routes.

Time Estimation: estimating rider travel time, indoor dwell time, restaurant preparation time, and expected delivery time.

Supply‑Demand Balancing: real‑time pressure balancing, medium‑ and long‑term order forecasting, capacity planning, and short‑term rider scheduling.

Location Services: defining delivery zones, building business‑district/building/delivery‑point maps, and calibrating locations.

Figure 1 Ele.me Intelligent Scheduling System

2. Pressure‑Balancing System

The pressure‑balancing system addresses situations where rider supply and order demand become mismatched within a day, aiming to protect user experience by applying timely control measures such as adjusting delivery fees, shrinking delivery ranges, launching discounts, or closing stores.

Figure 2 Pressure‑Balancing System Goal

The algorithm framework (Figure 3) is divided into three layers:

M1 – Real‑time Monitoring & Prediction: monitors team pressure coefficients and auxiliary indicators to determine when and how much to intervene.

M2 – Restaurant Understanding & Control‑Strategy Pool: ranks restaurants, predicts burst orders, assesses delivery difficulty, and stores possible control actions (e.g., fee increase, discount, range reduction).

M3 – Automatic Control‑Strategy Generation Model: optimizes a combination of metrics based on quantified control targets to decide which restaurants to adjust and which strategies to apply.

All modules are data‑driven and employ machine‑learning techniques; completed parts are highlighted in green, while ongoing or planned work appears in yellow.

Figure 3 Pressure‑Balancing Algorithm Framework

2.1 Data Monitoring

A real‑time monitoring dashboard (Figure 4) shows the normalized pressure coefficient for each team, updated every five minutes and adapting to weather, temperature, etc.

Figure 4 Pressure Coefficient Monitoring Dashboard

3. Machine‑Learning Algorithms in Pressure Balancing

3.1 Rider Maximum Order Capacity Model

The model estimates a rider's maximum simultaneous order capacity, a key component of rider profiling. Three versions were developed:

Version 1: simple rule‑based average capacity per team (too coarse).

Version 2: rule‑based tiers assigning the same capacity to riders within a tier (still coarse).

Version 3: a binary‑classification model predicting the probability of timeout for a given order load, using logistic regression.

Figure 5 Rider Capacity Model Iterations

Feature engineering includes:

Historical non‑timeout maximum order counts (averages over 28, 21, 14, 7, 5, 3 days).

Weather and temperature.

Rider personal attributes (level, tenure, team info).

Temporal features (day of week, workday flag).

Order‑specific features (maximum concurrent orders during delivery).

Logistic Regression was chosen for its simplicity, speed, interpretability, and ability to output probabilities via the sigmoid function. Training data comprised the most recent 14 days of orders during the lunch peak (10:30‑12:30). Because timeout cases are rare, non‑timeout samples were down‑sampled to ~10%.

The model is also used to compute the team pressure coefficient:

Team Pressure Coefficient = load / (q₁ + q₂ + … + qₙ)

where load is the team's current workload and qᵢ is the estimated maximum order capacity of rider i.

3.2 Team Pressure Coefficient Prediction Model

A real‑time predictor forecasts the pressure coefficient for the next 1.5 hours at 5‑minute intervals (six 15‑minute slices). The predictions guide two main scenarios:

Quantifying automatic control targets to trigger pre‑emptive adjustments.

Informing intelligent dispatch strategies based on anticipated pressure.

Three iterations of this model have been completed, steadily improving accuracy and prediction frequency (see Figure 6).

4. Summary and Outlook

The article aims to give readers insight into Ele.me's pressure‑balancing system within its instant‑delivery ecosystem and how common machine‑learning algorithms can effectively address O2O challenges. Future work will explore additional ML problems such as ranking learning and time‑series forecasting.

Recommended Reading

Machine Learning in Hotel Call‑Center Automation Building a React‑Based Mobile SPA Framework Step‑by‑Step Machine Learning in Web Attack Detection Distributed Architecture for Global Unique Sequence Generation Ctrip's Next‑Generation Call‑Center Monitoring Platform