💻 computer science

A Hybrid Multi-scale Convolutional Neural Network-Long Short-Term Memory Model Incorporating Squeeze Excitation Attention for Hourly Air Quality Index Prediction

This study proposes a hybrid MSCNN-LSTM model enhanced with Squeeze-and-Excitation attention to predict hourly Air Quality Index values, demonstrating superior accuracy and stability compared to traditional deep learning models through multi-scale feature extraction and adaptive channel weighting on data from Yiyang City.

Original authors: Huawei Xu, Long Li, Qi Tang, Junhua Huang, Xin Fang

Published 2026-09-08

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Huawei Xu, Long Li, Qi Tang, Junhua Huang, Xin Fang

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

Air pollution is a silent, shifting force that shapes the health of cities and the safety of the people who live in them. To manage this invisible threat, scientists rely on a single number called the Air Quality Index, or AQI. This index acts as a daily report card, translating complex mixtures of smoke, dust, and gases into a score that tells us whether the air is safe to breathe. However, the air does not stay still; it changes hour by hour, driven by the weather, traffic, and industrial activity. Predicting these changes is difficult because the factors involved interact in messy, non-linear ways. Traditional methods often struggle to see the full picture, missing the subtle patterns that signal a sudden spike in pollution or a rapid clearing of the sky. To solve this, researchers have turned to a branch of computer science known as deep learning, where computer programs learn to recognize patterns in vast amounts of data, much like a human brain learns from experience.

In a recent study focused on Yiyang City in Hunan Province, a team of researchers set out to build a better crystal ball for air quality. They wanted to create a system that could forecast the hourly AQI with high precision, using data from the past two years. The team, led by Huawei Xu and colleagues, developed a new hybrid computer model that combines three distinct digital tools to mimic the way the atmosphere behaves. First, they used a multi-scale convolutional neural network, a type of program designed to look at data through different "lenses" simultaneously. Imagine trying to understand a landscape by looking at it through a wide-angle lens, a telephoto lens, and a zoom lens all at once; this tool does the same for time, capturing short-term spikes, medium-term trends, and long-term cycles in the pollution data all at once.

Next, the team introduced a mechanism called Squeeze-and-Excitation attention. This acts as a smart filter for the information the computer has gathered. In a storm of data, not every piece of information is equally important at every moment. This attention mechanism allows the model to pause, evaluate which factors are currently driving the pollution levels, and amplify those signals while ignoring the noise. Finally, the system feeds this refined information into a Long Short-Term Memory network, or LSTM. This part of the model is specialized in remembering the past; it holds onto the history of how the air quality has changed over time, allowing it to understand how a cold morning might lead to a smoggy afternoon, or how a week of rain might clear the air. By weaving these three components together, the researchers created a model that is not just looking at the data, but understanding the story it tells.

The researchers tested their new model against older, simpler methods using real-world data from Yiyang City, covering every hour from January 2023 to December 2024. They fed the computer historical records of pollutants like fine particulate matter and carbon monoxide, alongside weather data such as temperature, wind, and atmospheric pressure. The results were striking. The new hybrid model outperformed the standard approaches, reducing its prediction errors by roughly a quarter compared to the basic models. It achieved a score of 0.96 on a scale where 1.0 is a perfect match, meaning its forecasts were incredibly close to what actually happened in the air. The model proved particularly adept at handling the extremes, accurately predicting both the dangerous peaks of winter pollution and the clean, low values of summer.

The study also revealed how the air quality in Yiyang behaves throughout the year. The data showed a clear rhythm: pollution levels are highest in the winter, often exceeding 100 on the index, driven by cold temperatures that trap smoke near the ground and the burning of fuels for heating. In contrast, summer air is significantly cleaner, with average scores dropping below 40, thanks to heavy rains that wash pollutants away and rising heat that helps disperse them. The computer model learned these seasonal rules naturally, without being explicitly told the calendar dates. When the researchers asked the model to explain which factors mattered most, it pointed to the same things a human expert would expect: the concentration of fine dust particles, the temperature, and the atmospheric pressure. The model confirmed that when these specific numbers rise or fall, the air quality follows suit.

By combining the ability to see patterns at different speeds, the intelligence to focus on what matters, and the memory to learn from history, this new approach offers a powerful tool for city planners and public health officials. It does not just guess the future; it calculates it based on the complex, shifting reality of the atmosphere. The study demonstrates that when we give computers the right tools to look at the world, they can help us see the invisible dangers in our air with unprecedented clarity, providing a reliable guide for protecting communities from the health risks of pollution.

Technical Summary: A Hybrid Multi-scale Convolutional Neural Network-Long Short-Term Memory Model Incorporating Squeeze Excitation Attention for Hourly Air Quality Index Prediction

Problem Statement
Accurate prediction of the Air Quality Index (AQI) is critical for public health protection and policy formulation. While existing deep learning models have been applied to AQI forecasting, they face specific limitations: standard Convolutional Neural Networks (CNNs) often rely on single-scale convolutional kernels, which fail to capture the complex multi-time-scale patterns inherent in air pollutant variations. Furthermore, traditional hybrid models (e.g., CNN-LSTM) frequently lack an adaptive mechanism to regulate the importance of feature channels when integrating multi-source data (pollutants and meteorology), leading to redundant information interfering with key feature representation. Additionally, existing models often struggle with prediction stability during low-value periods, such as summer months.

Methodology
To address these limitations, the authors propose a hybrid deep learning architecture: MSCNN-LSTM with SE Attention. The model integrates three core components:

Multi-Scale Convolutional Neural Network (MSCNN): Unlike standard CNNs, this module employs parallel convolutional layers with three distinct kernel sizes (3, 5, and 7) and varying filter counts (96, 64, and 32). This design extracts multi-scale temporal features from input sequences, capturing short-term fluctuations, mid-term trends, and long-term periodic patterns simultaneously.
Squeeze-and-Excitation (SE) Attention Mechanism: Following the concatenation of multi-scale features, the SE module performs a "Squeeze" operation (global average pooling) to compress spatiotemporal features into a channel descriptor. An "Excitation" operation then utilizes two fully connected layers (with a compression ratio of 3) to generate channel-specific weight coefficients. This allows the model to adaptively recalibrate feature channels, highlighting critical inputs (e.g., specific pollutants or meteorological factors) while suppressing redundant or noisy information.
Long Short-Term Memory (LSTM): The re-weighted feature sequences are fed into an LSTM layer to capture long-term temporal dependencies and cumulative effects in AQI variations, ultimately outputting the predicted AQI value.

Data and Experimental Setup
The study utilized hourly air pollutant and meteorological data from Yiyang City, Hunan Province, covering the period from January 1, 2023, to December 31, 2024. Input variables included PM2.5, PM10, NO2, CO, temperature, dew point, atmospheric pressure, and periodic time features (encoded via sine/cosine functions). The dataset was split into 80% for training and 20% for validation. The proposed model was benchmarked against three other architectures: standard LSTM, LSTM with SE Attention, and a standard CNN-LSTM. Performance was evaluated using Mean Absolute Error (MAE), Root Mean Square Error (RMSE), Mean Absolute Percentage Error (MAPE), and the Coefficient of Determination ( $R^2$ ).

Key Results

Overall Performance: The MSCNN-LSTM with SE Attention model achieved the highest predictive accuracy on the validation set, with an $R^2$ of 0.96. Compared to the traditional LSTM model, it reduced MAE by 25.16%, RMSE by 22.14%, and MAPE by 23.17%, with values of 5.65, 7.94, and 8.72%, respectively.
Seasonal Stability: The proposed model outperformed the standard LSTM across all four seasons. Notably, it demonstrated superior fitting capability and stability during summer months (low AQI periods), a scenario where other models often exhibited higher errors. In summer scenarios, the model reduced MAE, RMSE, and MAPE by 5%, 5%, and 9% respectively compared to the CNN-LSTM model.
Feature Importance: SHAP (SHapley Additive exPlanations) analysis identified PM2.5, PM10, temperature, and atmospheric pressure as the most significant factors influencing AQI predictions. The analysis confirmed that the model's predictions align with physical mechanisms, such as the positive correlation between high temperatures/pressure and increased AQI due to reduced dispersion and enhanced photochemical reactions.

Significance and Claims
The paper claims that the proposed hybrid model provides a high-precision and interpretable solution for hourly regional air quality forecasting. Its primary significance lies in its ability to simultaneously address two critical gaps in existing literature: the inability of single-scale kernels to capture multi-temporal variations and the lack of adaptive channel weighting in multi-source feature integration. By effectively combining multi-scale feature extraction with dynamic attention mechanisms, the model achieves superior robustness against extreme AQI values and improved stability during low-pollution periods. The authors conclude that this architecture offers a reliable tool for pollution control, early warning systems, and the scientific basis for regional air quality policy formulation.

More like this