Medium and Short-Term Wind Power Forecasting Using LCNN-Mamba Model Based on Quadratic Decomposition
This paper proposes a hybrid LCNN-Mamba model enhanced by quadratic decomposition (STL and CEEMDAN) to effectively address the non-stationarity and multiscale fluctuations in wind power data, achieving superior short-to-medium-term forecasting accuracy through the synergistic integration of local convolutional feature extraction and global state-space dependency modeling.
Original paper licensed under CC BY 4.0 (https://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 trying to predict the weather, but instead of clouds and rain, you're tracking the invisible, fickle breath of the wind that spins giant turbines to make electricity. This is the world of wind power forecasting, a critical puzzle for scientists and engineers. Why does it matter? Because wind is a bit of a mood ring; it blows hard one minute and goes calm the next. If power grid operators can't guess how much energy the wind will produce, they might have to burn fossil fuels as a backup or risk blackouts. To solve this, researchers use "time series forecasting," which is just a fancy way of saying "looking at a line of data points from the past to guess what the next point will be." But wind data is messy, jumping around wildly and changing patterns constantly. To make sense of this chaos, scientists often use "decomposition," a technique like taking a complex song and separating it into the bass line, the melody, and the drums so they can study each part individually. They also use "deep learning," which are computer programs inspired by the human brain that get really good at spotting patterns in huge amounts of data. The goal is to build a system that can look at the wind's chaotic history and say, "Okay, I know what's coming next," so our lights stay on and our planet stays clean.
Now, meet the team from Inner Mongolia University of Technology who decided to build a super-smart crystal ball for the wind. They realized that existing tools were like trying to catch a greased pig: some models were great at seeing the big picture but missed the sudden jumps, while others were good at the small details but got lost in the long-term trends. So, they cooked up a new recipe called the LCNN-Mamba Model, which is built on a "quadratic decomposition" strategy. Think of this strategy as a two-step cleaning process for a dirty window. First, they use a method called STL to wipe away the long-term smudges (the slow, steady trends) and the regular streaks (the daily or seasonal cycles). Once the window is mostly clean, they take the remaining "residue"—the tricky, random splatters—and run it through a second, more intense cleaning machine called CEEMDAN. This second step breaks the messy leftovers into tiny, manageable pieces based on how fast they vibrate (high, medium, and low frequencies).
But cleaning the data is only half the battle; you still need a brain to read it. That's where their new LCNN-Mamba model comes in. Imagine a detective team where one partner is a Local Convolutional Neural Network (LCNN). This partner is hyper-focused on the immediate neighborhood, spotting sudden, sharp changes like a gust of wind that spikes the power output in a split second. The other partner is Mamba, a state-of-the-art AI that acts like a historian, looking way back into the past to understand the long-term story and slow-moving trends. By teaming these two up, the model gets the best of both worlds: it notices the sudden "oops" moments without getting confused by the long-term story, and it understands the big picture without missing the tiny details.
The researchers tested this new system using real data from a wind farm in Inner Mongolia, covering over 76,000 data points collected every 15 minutes between January 2021 and March 2023. They compared their new model against older, famous models like LSTM, GRU, and even the popular Transformer. The results were clear: their new hybrid model was the star of the show. In a test predicting 192 steps ahead (which is about 48 hours of wind data), their model achieved an accuracy score (R²) of 0.996, meaning it was almost perfectly aligned with reality. In comparison, older models like LSTM dropped to 0.889 for the same task. The new model also kept its errors incredibly small, with an average mistake (MAE) of just 0.418 MW, whereas the older models made mistakes nearly twice as big.
To prove that every part of their machine was necessary, they ran "ablation experiments," which is like taking a car apart to see which engine part actually makes it go. When they removed the first cleaning step (STL), the model's accuracy dropped significantly. When they removed the second cleaning step (CEEMDAN), the errors got messy and wide. When they took out the "Local" detective (LCNN), the model smoothed over important spikes and missed the sudden changes. And when they removed the "Historian" (Mamba), the model lost its ability to track the long-term trends, resulting in wild, unstable predictions. This confirmed that the magic wasn't in just one part, but in how all the pieces—the two-stage cleaning, the local detector, and the long-term historian—worked together.
They didn't stop there. To see if their model was truly tough, they tried it on data from a completely different wind farm in Ningxia, over 500 kilometers away, with different weather and terrain. Even with this new, unfamiliar data, the model held up well, maintaining a high accuracy of 0.996 for short-term predictions and 0.908 for longer ones. This suggests the model isn't just memorizing one specific wind farm; it's actually learning how to understand wind in general.
In the end, the paper suggests that by breaking wind data down into its simplest parts and using a team of specialized AI detectives to read them, we can predict wind power with much higher precision. This isn't just a math trick; it's a practical tool that could help power grids run smoother, reduce waste, and help us rely more on clean energy. While the authors note that more testing is needed to see how it works in other parts of the world, their current findings show a promising path toward a more reliable, wind-powered future.
Drowning in papers in your field?
Get daily digests of the most novel papers matching your research keywords — with technical summaries, in your language.