Data-Driven Mission Recommendation for LEO Satellites using Random Forest and Hybrid Deep Learning Forecasting

Imagine the sky above us is filled with tiny, high-speed robots called satellites. These aren't just floating cameras; they are complex machines that need to eat energy to do their jobs, like taking photos or sending signals. They get their food from the sun, but the sun doesn't shine on them all the time. When they fly into the Earth's shadow, they have to rely on a battery, much like your phone does when you forget your charger. The big challenge for the people on the ground who control these robots is knowing exactly how much "battery juice" is left before they ask the satellite to do a new, energy-hungry task. If they ask for too much, the satellite might pass out and fail its mission.

For a long time, figuring this out was like trying to guess the weather by looking at the clouds and hoping for the best. Experienced operators would use their gut feelings and years of memory to decide if a satellite had enough power. But memories can be fuzzy, and different people might make different guesses. This paper introduces a new way to solve that puzzle using a "digital brain" that learns from six years of history. It combines two types of smart math: one that spots repeating patterns (like the sun rising and setting) and another that learns from the messy, leftover details. The goal is to turn a vague feeling into a clear, written list of exactly which tasks the satellite can safely do right now.

---

The Satellite's Energy Dilemma

Meet LAPAN-IPB, a satellite orbiting 500 kilometers above Indonesia. It's a polar satellite, meaning it zooms over the North and South poles, circling the Earth about 15 times a day. Every time it flies over Indonesia, it gets a short window—about 10 to 15 minutes—to talk to ground stations. During this tiny window, the operators on the ground can send it new instructions, like "take a picture of this forest" or "scan the ocean."

But here's the catch: the satellite is running on a battery. Between those short windows, it spends time in the dark (eclipse), living off its stored energy. If the battery is too low, and the operators send a command that uses too much power, the satellite could crash or lose its ability to function.

For years, the operators at the National Research and Innovation Agency (BRIN) in Indonesia had to make these life-or-death decisions based on "tacit knowledge." That's a fancy way of saying "gut feeling built on experience." If an operator had been there for ten years, they might know, "Oh, the battery looks a bit low today, so let's skip the heavy camera mission." But if a new operator took over, they wouldn't have that experience. They might guess wrong, or two operators might argue about the same situation. There was no written rulebook, no clear audit trail, and no way to prove why a decision was made.

The New "Digital Co-Pilot"

This paper describes a new system designed to replace that guesswork with a data-driven recommendation engine. Think of it as a co-pilot that reads the satellite's vital signs and hands the operator a ranked list of safe tasks. The system was built using six years of data (from January 2019 to December 2024), which included over 25,000 records of the satellite's battery voltage and current.

The researchers broke the problem down into three main steps, using a clever mix of old-school statistics and modern deep learning.

1. Predicting the Future Voltage
First, the system needs to guess what the battery voltage will be at the next time the satellite talks to the ground. To do this, they used a "Hybrid" model. Imagine trying to predict the tide. You know the tide goes up and down in a regular cycle (that's the easy part). But sometimes, a storm or a strong wind changes the water level unexpectedly (that's the hard part).

The researchers used a statistical model called SARIMAX to handle the regular, predictable up-and-down cycle of the battery charging and discharging. Then, they used a Deep Learning model (specifically a type called LSTM) to figure out the messy, unpredictable leftovers. By combining them, they got a very accurate prediction. On their test data, this hybrid model was able to predict the voltage with an average error of just 0.1042 V (Volts). That's incredibly precise, especially since the battery voltage usually hovers around 16 to 17 Volts.

2. Learning the "Energy Fingerprints"
Next, the system needed to know how much energy each specific mission uses. The satellite has 20 different types of missions (like "take a photo," "scan the ocean," etc.), and each one turns on different switches and uses different amounts of power.

The researchers taught a Random Forest classifier (a type of machine learning that works like a team of decision trees) to recognize the "energy fingerprint" of each mission. They looked at the battery current and voltage during past missions to see what each one "felt" like. However, they had to be careful. The satellite sends data many times during a single 15-minute pass. If they treated every single data point as a separate lesson, the computer would get confused, thinking it was learning from 25,000 different events when it was actually just seeing the same event repeated over and over. They fixed this by grouping the data into "passes," so the computer learned from 6,833 unique events instead of 25,000 confusing duplicates.

The result? The computer learned to identify the 20 mission types with a high degree of accuracy, correctly guessing the mission type about 82% of the time, and doing a great job even with the rare, less common missions.

3. The Safety Check and Ranking
Finally, the system brings it all together. It takes the predicted future voltage and compares it to a "safety threshold" for each of the 20 missions. Where did these thresholds come from? The researchers used a clever statistical trick called the Tukey upper fence. Imagine you have a pile of data showing how high the battery voltage was when a specific mission was successfully run in the past. The Tukey fence finds the "safe upper limit" of that pile. If the predicted voltage is higher than this limit, the mission is safe.

To make sure these limits were trustworthy, they checked them in two ways: once using the raw data, and once using the internal logic of the machine learning model. They found that for all 20 missions, the two methods agreed almost perfectly (the difference was less than 0.0843 V, which is smaller than the prediction error itself). This means the safety limits are rock-solid.

The Result: A Clear, Ranked List

When the satellite is about to make contact with the ground, the system runs a quick calculation. It predicts the battery voltage, checks it against the 20 safety limits, and spits out a ranked list.

• Feasible Missions: These are the tasks the battery can definitely handle.
• The Ranking: The system doesn't just say "Yes" or "No." It ranks the "Yes" missions by their energy margin. This is the difference between the predicted voltage and the safety limit.
  • A mission with a huge margin (e.g., +0.77 V) is super safe.
  • A mission with a tiny margin (e.g., +0.02 V) is technically possible but risky, like walking a tightrope.

This gives the operator a clear, evidence-based starting point. Instead of guessing, they can see: "Mission A is safe with a huge buffer, but Mission B is cutting it close." If the battery is low, the system might say, "Only these three low-power tasks are safe today." If the battery is full, it might say, "You can do almost anything, but Mission C is the safest bet."

Why This Matters

The biggest win here isn't just that the math works; it's that the process is now reproducible and auditable. Before, if a mission failed, you couldn't be sure if it was because the operator made a bad guess or if the satellite was just unlucky. Now, every decision is backed by six years of data and a clear, documented rule.

The system doesn't replace the human operator; it gives them a superpower. It turns a vague feeling into a precise, ranked list of options, ensuring that no matter who is on shift, the satellite gets the best possible care. The researchers found that this approach is stable, reliable, and ready to help the LAPAN-IPB satellite keep flying safely for years to come.