Energy Efficient Traffic Scheduling For Optical LEO Satellite Downlinks

This paper proposes and evaluates static and adaptive traffic scheduling schemes—including threshold, heuristic, and reinforcement learning-based approaches—to optimize energy efficiency and delivery ratios for energy-constrained optical LEO satellite downlinks facing weather-related disruptions.

The Gist

Imagine a fleet of satellites zooming around the Earth like high-speed delivery drones. Their job is to pick up data (like photos from Earth observation or messages from remote sensors) and drop it off at ground stations.

Now, imagine these satellites are trying to drop off packages using a laser beam instead of a radio wave. This is called Free-Space Optical (FSO) communication. It's incredibly fast, like a fiber-optic cable in the sky. But there's a catch: clouds.

If a cloud passes between the satellite and the ground station, the laser beam gets blocked. The satellite still has to spend energy to aim, lock on, and try to transmit, but the data bounces off the cloud and is lost. It's like trying to shout a message to a friend across a valley, but a thick fog rolls in. You're still screaming (wasting energy), but they can't hear you.

Since satellites run on batteries (solar power), wasting energy on failed attempts is a big problem. The goal of this paper is to figure out when the satellite should try to send data and when it should just wait, so it doesn't waste its battery.

The Big Problem: The "Cloudy Day" Dilemma

The researchers looked at two main types of traffic:

1. Urgent Traffic: Like a phone call or a video stream. This needs to happen now.
2. Patient Traffic: Like Earth photos or sensor data. This can wait hours or even days.

The paper focuses on the Patient Traffic. Because it can wait, the satellite doesn't need to force a transmission the second it sees the ground. It can be smart about it.

The Solution: A Smart Scheduler

The researchers created different "strategies" (algorithms) to decide which "contact windows" (times when the satellite is visible from the ground) are worth using. They tested four main types of strategies:

1. The "Always Try" Strategy (Baseline)

This is the old-school way. The satellite sees the ground, and it tries to send data immediately, no matter what the weather looks like.

• Analogy: It's like a mailman who tries to deliver a letter to every house on the street, even if it's pouring rain and the porch is flooded. He gets wet (wastes energy), and the letter might get ruined.
• Result: He delivers almost everything (high delivery ratio), but he gets soaked and tired (low energy efficiency).

2. The "Cloud Threshold" Strategy (Static)

This strategy sets a simple rule: "Only try to send if the cloud cover is less than 50%."

• Analogy: The mailman checks the sky. If it's sunny or partly cloudy, he delivers. If it's stormy, he goes home.
• Result: He saves a lot of energy. But if the storm is short and he misses a perfect 10-minute window, he might leave a letter behind.
• The Flaw: The rule is rigid. If the weather changes unexpectedly, the mailman can't adapt. He might miss a sunny break because his rule was too strict.

3. The "Smart Sorter" Strategy (Static & Adaptive)

Instead of a simple rule, this strategy looks at the entire schedule for the day. It ranks the contact windows from "Best Weather" to "Worst Weather" and picks the best ones first.

• Analogy: The mailman looks at the whole week's forecast. He plans to deliver to the sunniest houses first.
• The Adaptive Twist: The "Adaptive" version is even smarter. After he tries to deliver to one house, he re-checks the weather for the rest of the day. If the clouds move, he changes his plan instantly.
• Result: This is the most efficient. It saves the most energy while still getting most of the letters delivered.

4. The "AI Learner" Strategy (Reinforcement Learning)

This uses Artificial Intelligence (AI) to learn from experience. The satellite acts like a student: it tries a strategy, sees if it worked, and learns from its mistakes to get better next time.

• Analogy: The mailman doesn't just follow a map; he learns from his own experience. "Oh, I tried to deliver at 2 PM on Tuesdays and it always rained. Next Tuesday, I'll wait until 4 PM."
• Result: In a perfect, predictable world, this AI is amazing. But in the real world, where weather forecasts aren't 100% accurate, the AI sometimes gets confused and makes worse decisions than the simple "Smart Sorter."

What Did They Find?

The researchers ran thousands of simulations, including one using real historical weather data from Canadian cities. Here is the takeaway:

• Adaptive is King (mostly): The strategies that can change their minds in real-time (Adaptive Sorting) performed the best. They balanced saving energy with delivering data perfectly.
• Static is Simple but Rigid: The simple "Cloud Threshold" rules were easy to run on the satellite's computer, but they failed when the weather was unpredictable.
• AI is Tricky: The AI (Reinforcement Learning) was very powerful in simulations, but when they added real-world "noise" (imperfect weather forecasts), it struggled. It's like a student who studied for a specific test but panicked when the questions were slightly different.
• The Trade-off: The best-performing strategies require more computing power. If the satellite is a tiny, cheap CubeSat with a weak computer, it might not be able to run the "Smart AI." It might have to settle for the simpler "Cloud Threshold" rule.

The Bottom Line

For satellites sending data that can wait, being flexible is key.

If you have a powerful computer on your satellite, use an Adaptive Strategy that constantly re-evaluates the weather. It will save you the most battery life while ensuring your data gets home. If your satellite is small and simple, a Static Strategy is better than just blindly trying to send data, but you have to be careful not to set the rules too strictly, or you might miss your chance to deliver.

In short: Don't scream into the fog. Wait for a break in the clouds, and if you're smart enough, change your plan the moment the clouds move.

Technical Summary

1. Problem Statement

The paper addresses the challenge of optimizing Free-Space Optical (FSO) downlinks for Low Earth Orbit (LEO) satellites carrying delay-tolerant traffic (e.g., Earth Observation data, Satellite IoT).

• Context: While megaconstellations (like Starlink) handle delay-sensitive traffic, sparse constellations are more cost-effective for delay-tolerant services. However, these networks face a bottleneck in space-to-ground link capacity.
• The Challenge: FSO links offer high data rates but are highly susceptible to weather, specifically cloud cover, which can cause link failure or severe attenuation.
• The Core Conflict:
  • Energy Constraints: Satellites have limited power. Establishing an FSO link requires energy for pointing, acquisition, and tracking. Attempting transmission during poor weather (cloud cover) wastes this energy if the link fails.
  • Delivery Requirements: Despite being "delay-tolerant," there is a need to maximize the Delivery Ratio (DR) (amount of data successfully sent) while minimizing Energy Expenditure.
• The Gap: Existing scheduling works focus on delay-sensitive networks or assume deterministic links. There is a lack of energy-efficient scheduling strategies specifically for delay-tolerant networks that account for the stochastic nature of weather-induced link failures.

2. Methodology

A. System Model & Problem Formulation

• Architecture: A single LEO satellite communicating with a single ground station. The system operates in "downlink periods" where a batch of data must be delivered.
• Link Modeling: The channel is modeled as an on-off stochastic process. Link availability ($CA_i$) is determined by cloud cover ($CC_i$) and contact volume ($v_i$).
  • Energy Waste: Defined as the energy spent on unsuccessful packet transmissions due to cloud cover.
• Mathematical Formulation: The authors formulate the scheduling problem as a variant of the 0-1 Knapsack Problem.
  • Items: Transmission contacts.
  • Value: A function of delivered packets and energy efficiency.
  • Weight: The data volume required.
  • Constraint: The total data to be delivered.
  • Objective: Maximize a weighted sum of Delivery Ratio and Energy Efficiency.

B. Proposed Scheduling Schemes

The authors propose and evaluate two categories of schemes: Static (pre-determined logic) and Adaptive (logic adjusts based on real-time state).

1. Static Schemes:

• Baseline (CGR): Uses Contact Graph Routing to utilize all available contacts until data is exhausted. Maximizes DR but ignores energy efficiency.
• Threshold Schemes: Only attempts transmission if the predicted cloud cover is below a pre-calculated threshold ($T$). Includes a tuning algorithm (Algorithm 3) to find optimal thresholds based on historical data.
• Static Sorting: Sorts all available contacts by predicted cloud cover (best to worst) and selects the top contacts needed to meet the data volume, ignoring real-time changes during the downlink.

2. Adaptive Schemes:

• Adaptive Sorting: Similar to static sorting but re-evaluates and re-sorts the remaining contacts after every transmission attempt, accounting for the actual data delivered and remaining volume.
• Reinforcement Learning (RL) & Deep RL (DRL):
  • Formulated as a Markov Decision Process (MDP).
  • State: Cloud cover, remaining data volume, remaining system capacity, and future contact info.
  • Action: Binary decision (Transmit = 1, Wait = 0).
  • Reward: A composite function rewarding successful delivery and penalizing energy waste (failed transmissions).
  • Algorithms: Q-learning (tabular) and Double Deep Q-Network (DDQN) (neural network-based) were implemented to learn optimal policies.

C. Complexity Analysis

• Static (CGR/Threshold): $O(N)$ complexity.
• Static Sorting: $O(N \log N)$ due to sorting.
• Adaptive Sorting: $O(N^2 \log N)$ due to re-sorting at every step.
• RL/DRL: $O(N^2)$ inference time complexity.
• Note: Static schemes are computationally lighter, making them more suitable for resource-constrained satellites.

3. Key Contributions

1. Problem Formulation: Successfully mapped the optical downlink scheduling problem for delay-tolerant traffic to a Knapsack Problem variant, explicitly modeling the trade-off between delivery ratio and energy waste.
2. Scheme Development: Developed a suite of static and adaptive scheduling algorithms (Threshold, Sorting, RL, DRL) tailored for sparse LEO networks with weather constraints.
3. Performance Validation: Demonstrated that adaptive schemes significantly outperform static schemes in dynamic environments (changing weather/data volumes) but at the cost of higher computational complexity.
4. Real-World Case Study: Validated algorithms using historical weather data (2023-2024) and realistic orbital parameters for Canadian ground stations, highlighting the gap between ideal simulations and real-world uncertainty.

4. Key Results

A. Uniform & Controlled Simulations

• Delivery Ratio: Adaptive schemes (DDQN, Adaptive Sorting) maintained near-perfect delivery ratios even when conditions changed, whereas static schemes (especially Threshold) suffered significant drops (up to 7-13% reduction) when data volume or cloud cover distributions shifted.
• Energy Efficiency:
  • Adaptive Sorting achieved the highest energy efficiency improvements (~24.7% better than baseline CGR).
  • DDQN showed strong efficiency (~14.8% improvement) with minimal loss in delivery ratio.
  • Static Thresholds: While efficient in tuned scenarios, they failed to maintain delivery ratios when conditions deviated from the training data.
• Q-learning: Performed poorly compared to DDQN, suggesting that tabular methods struggle with the continuous state space of this problem without heavy discretization.

B. Case Study (Real-World Data)

• Impact of Uncertainty: When tested with realistic weather forecast uncertainty and variable ground station locations:
  • DDQN performance degraded significantly, performing worse than simple threshold schemes. This indicates that current DRL models are sensitive to the gap between training simulation and real-world channel dynamics.
  • Sorting Algorithms (Static and Adaptive) proved more robust than RL in this specific case study, though Adaptive Sorting still suffered some delivery ratio degradation compared to the baseline.
• Trade-off: The study confirmed that while adaptive schemes offer superior theoretical performance, their reliance on accurate channel models and higher onboard computation makes them less feasible for current hardware compared to simpler static schemes in highly uncertain environments.

5. Significance and Conclusion

• Energy Efficiency: The paper proves that intelligent scheduling can significantly reduce the energy wasted on failed optical links, a critical factor for extending satellite mission life.
• Adaptability vs. Complexity: The central finding is the trade-off between adaptability and computational cost.
  • Static schemes are robust, low-complexity, and suitable for current hardware but fail in highly dynamic scenarios.
  • Adaptive schemes (especially Sorting) offer the best balance of performance and robustness for delay-tolerant traffic, whereas DRL currently struggles with the "reality gap" of weather prediction uncertainty.
• Future Directions: The authors suggest that future work should focus on Offline RL or Online Learning to handle channel dynamics without perfect prior knowledge, and extending the model to multi-satellite constellations with limited ground terminals.

In summary, this work provides a comprehensive framework for optimizing optical downlinks in delay-tolerant satellite networks, demonstrating that while advanced AI (DRL) is promising, adaptive heuristic sorting currently offers the most practical balance of energy efficiency and delivery reliability for near-future satellite deployments.