Leakage-Safe K-Means-Assisted Multi-Objective Energy Management for a Solar and Wind–Battery Micro-grid

Imagine a small, self-contained power system that runs a community or a remote facility without relying on the main electrical grid. This system, known as a micro-grid, tries to balance three moving parts: electricity generated by the sun and wind, electricity stored in large batteries, and electricity produced by a backup generator. The challenge is that the sun does not always shine, the wind does not always blow, and people do not always use power at the same time. To keep the lights on without wasting fuel or draining the batteries, a computer system must constantly decide when to charge the batteries, when to run the generator, and when to save the renewable energy for later. This decision-making process is called energy management. If the system guesses wrong, it might run out of power, or it might waste expensive fuel by running the generator when it could have used the sun.

Researchers at Kocaeli University in Turkey set out to build a smarter way to make these decisions. They wanted to create a system that could look at the past and predict the immediate future to manage a solar and wind-powered micro-grid equipped with a battery. Their approach involved two main ideas. First, they used a computer method called Random Forest, which is a type of machine learning that learns from past data to predict what the power demand will be one hour from now. Second, they used a technique called K-Means clustering to group similar weather and usage patterns together. Think of these groups as different "modes" the system might find itself in, such as a sunny day with low power use or a windy night with high demand. The goal was to see if knowing which "mode" the system was in could help the computer make better choices about when to use the battery and when to start the generator.

The researchers were very careful to avoid a common mistake in computer science called data leakage. This happens when a computer learns from the future while trying to predict the present, which makes the results look perfect but useless in the real world. To prevent this, they split their data into three distinct time periods that never overlapped. They used the first period to teach the computer how to predict power needs and how to identify the different operating modes. They used the second period to tune the rules the computer would follow, ensuring the system would not run out of power or drain the battery too much. Finally, they locked the system's settings and tested it on the third period, which the computer had never seen before. This strict separation ensured that the results reflected how the system would actually perform in the real world, not just how well it memorized past data.

The results of this careful testing showed that the computer's ability to predict power demand was quite good. Using the Random Forest method, the system could guess the power needs for the next hour with an average error of about 2.3 kilowatts, which was significantly better than simply assuming the next hour would be exactly like the current one. However, the study found something surprising about the second part of their plan. Adding the "mode" groups from the K-Means clustering did not actually help the computer predict the power demand any better. The groups were just as accurate without the extra categorization. Because of this, the researchers decided not to use the groups to improve the prediction itself. Instead, they used the groups to help the system choose the right strategy for handling power. For example, if the system recognized a "sunny surplus" mode, it would know to charge the batteries; if it saw a "low power" mode, it would know to be ready to use the generator.

When they tested this new approach against older methods, the results showed a clear trade-off. An older system that reacted only to what was happening right now used the battery heavily, draining it quickly to avoid using the generator. Another system that relied only on predictions was very careful with the battery but ended up using the generator much more often and wasting a lot of renewable energy that could have been stored. The new system, which used the predicted demand and the operating modes to make decisions, demonstrated specific operational trade-offs. It reduced the amount of fuel the generator needed by nearly 60 percent compared to the prediction-only system, while also cutting down on wasted renewable energy by almost 24 percent. At the same time, it used the battery less than the reactive system, extending the battery's life. However, this improvement came with a cost: compared to the reactive baseline, the new system actually increased generator energy use by nearly 5,000 kWh and increased renewable energy curtailment by nearly 6,000 kWh. Despite this increase in generator use and waste compared to the reactive approach, the new system kept the battery at a safe level of charge almost all the time, avoiding the dangerous situation where the battery runs too low.

However, a critical detail emerged during the testing phase: the method used to select the best system settings did not actually identify a single, unique "best" solution. When the researchers tested 81 different feasible combinations of settings, they all produced identical performance scores. Because the test data could not distinguish between these options, the researchers reported the final system not as a uniquely optimized solution, but as a "feasibility-screened representative policy." This means the system is a valid, working example that meets all safety and reliability requirements, rather than the single mathematically perfect answer. The study concluded that this approach works well within the simulated environment they created. The system successfully managed to keep the power supply reliable, with zero instances of power failure during the test period. It achieved this by making smarter choices about when to store energy and when to generate it, rather than just reacting to the moment or relying on a single prediction. The researchers noted that while the results are promising, they are based on a simulation of a specific type of battery and a single year of weather data. They emphasized that the method's strength lies in its strict testing rules, which proved that the system could learn from the past without looking at the future. This work provides a solid blueprint for how to build energy management systems that are both efficient and reliable, provided they are tested with the same level of care before being installed in the real world.