A prediction model for hybrid air pollutants combining two-stage decomposition and two-scale feature selection

Air pollution is a constant presence in modern life, but one specific type of invisible threat often goes unnoticed until it causes harm: ground-level ozone. Unlike the protective layer high in the atmosphere, ozone near the ground is a toxic byproduct of sunlight reacting with other pollutants. It is a primary driver of smog and a significant health risk, particularly for children and the elderly, often triggering respiratory issues and damaging crops. Because ozone levels can shift rapidly and unpredictably, knowing when concentrations will spike is vital for public health. It allows people to plan their day, helps doctors prepare for an influx of patients, and gives city planners the data needed to manage traffic and industry. However, predicting these spikes is notoriously difficult because the data is messy, filled with sudden jumps and hidden patterns that standard forecasting tools struggle to decode.

A team of researchers from PowerChina Guizhou Electric Power Engineering Co., Ltd and Fuzhou University has developed a new approach to solve this problem. They created a prediction model designed specifically to handle the chaotic nature of ozone data. Instead of trying to predict the raw, jagged numbers directly, their method breaks the data down into smaller, simpler pieces, cleans up the noise, and then rebuilds the forecast with greater precision. By testing their system on daily ozone records from Beijing and Shanghai, the team found that their new model significantly outperformed existing methods, offering a clearer, more reliable picture of when dangerous air quality might occur.

The core of their strategy involves a process of breaking things down to understand them better. Imagine trying to listen to a conversation in a crowded room; if you focus on the whole noise, you hear nothing. But if you isolate the specific voices, the message becomes clear. The researchers applied a similar logic to the ozone data. They first used a technique called improved complete adaptive noise ensemble empirical mode decomposition to split the complex, volatile history of ozone levels into several distinct layers. This initial step smoothed out the erratic spikes, turning a chaotic signal into a set of more regular, manageable patterns.

However, the researchers noticed that even after this first split, the very first layer of data remained too complicated to predict accurately. It was still too noisy and unpredictable. To address this, they applied a second, more focused decomposition to just that one difficult layer, breaking it down further until it became simple enough to analyze. They then grouped the remaining layers based on how complex they were, merging those with similar characteristics into single, cleaner streams. This two-stage approach ensured that every piece of data fed into the final prediction was as simple and clear as possible.

Once the data was cleaned and simplified, the team faced another challenge: deciding which parts of the past history actually mattered for the future. In machine learning, feeding a model too much information can confuse it, a problem known as the "curse of dimensionality." To avoid this, the researchers used a two-step filter. First, they identified which past days had a statistical connection to the current day. Then, they used a computer algorithm to rank these connections and select only the most important ones, discarding the rest. This ensured the model was not overwhelmed by irrelevant details, allowing it to focus on the true drivers of ozone changes.

Finally, the cleaned and filtered data was fed into a specialized type of artificial intelligence called a temporal convolutional network. This system is designed to look back over time, capturing long-term patterns that other models might miss. By combining this powerful predictor with their rigorous data preparation steps, the researchers achieved remarkable results. In tests conducted in Beijing and Shanghai, their model produced errors that were significantly smaller than those of other popular forecasting methods. Specifically, the new model achieved a mean absolute error of just 0.53509 and a root mean square error of 0.73346 in Beijing, with similar high accuracy in Shanghai. These numbers indicate that the model's predictions were extremely close to the actual observed values, far surpassing the performance of standard models that did not use this two-stage decomposition and feature selection process.

The study demonstrates that the key to predicting complex environmental phenomena lies not just in having a powerful computer, but in how the data is prepared before the computer ever sees it. By systematically stripping away noise and focusing only on the most relevant historical patterns, the researchers created a tool that is both robust and precise. While the current model focuses solely on ozone levels, the authors suggest that future work could expand to include other factors like wind speed and temperature, potentially making the system even more effective. For now, this new approach offers a valuable reference for air quality control, providing a more intuitive and reliable way to anticipate the invisible threats in the air we breathe.