Source-linked AI summary
Deep Air Quality Forecasting Using Hybrid Deep Learning Framework
Shengdong Du, Tianrui Li, Yan Yang, Shi-Jinn Horng
TL;DR
Air-quality forecasting must handle dynamic, nonlinear, and interdependent multivariate time-series data relevant to pollution warning and control. The paper proposes DAQFF, a hybrid framework combining 1D-CNNs and Bi-LSTM to learn local trends, spatial correlations, and spatial-temporal dependencies. Tests on two real-world datasets report better single-step and multi-step forecasting ability than typical shallow and baseline deep-learning models.
Problem
Dynamic, nonlinear, and interdependent multivariate air-quality time series make accurate PM2.5 forecasting difficult, although forecasting supports pollution warning and control.
Method
DAQFF combines multiple 1D-CNNs and Bi-directional LSTM networks to learn local trend, spatial-correlation, and spatial-temporal dependency features from multivariate air-quality time series.
Results
DAQFF showed better single-step and multi-step forecasting ability than typical shallow-learning and baseline deep-learning models on two real-world air-quality datasets.
Takeaways & Limitations
The framework can effectively learn interdependence and nonlinear correlations among air-quality variables for PM2.5 forecasting tasks.
Takeaways & Limitations
Abrupt changes or anomalies in air-pollution time series require further study, and the model needs improvement under different forecasting conditions.
Abstract
from arXiv · showhide
Air quality forecasting has been regarded as the key problem of air pollution early warning and control management. In this paper, we propose a novel deep learning model for air quality (mainly PM2.5) forecasting, which learns the spatial-temporal correlation features and interdependence of multivariate air quality related time series data by hybrid deep learning architecture. Due to the nonlinear and dynamic characteristics of multivariate air quality time series data, the base modules of our model include one-dimensional Convolutional Neural Networks (1D-CNNs) and Bi-directional Long Short-term Memory networks (Bi-LSTM). The former is to extract the local trend features and spatial correlation features, and the latter is to learn spatial-temporal dependencies. Then we design a jointly hybrid deep learning framework based on one-dimensional CNNs and Bi-LSTM for shared representation features learning of multivariate air quality related time series data. We conduct extensive experimental evaluations using two real-world datasets, and the results show that our model is capable of dealing with PM2.5 air pollution forecasting with satisfied accuracy.
1 INTRODUCTION
Air quality forecasting is important for pollution warning and control, but traditional statistical and shallow-learning approaches struggle with dynamic, nonlinear, multivariate data. DAQFF addresses these characteristics with a hybrid deep-learning framework and reports effective forecasting performance.
- Air quality forecasting supports early pollution diagnosis, severe-pollution warning, and control decisions, particularly for PM2.5 concentration evolution.
- Weather, traffic, and other factors make air pollution difficult to represent and predict accurately using statistical methods and shallow machine-learning models.
- Growing sensor-based collections of PM2.5, NO2, PM10, weather, and traffic data motivate data-driven methods for multivariate air-quality time-series forecasting.
- DAQFF is an end-to-end hybrid deep-learning framework designed to model dynamic, spatial-temporal, and nonlinear characteristics in multivariate air-quality time series.
- The proposed model learns local trends and long spatial-temporal dependencies, and experiments report good forecasting performance and generalization ability.
- The paper evaluates the framework through comparative experiments analyzing its effectiveness for air-quality forecasting.
2 RELATED WORKS
Prior air-quality forecasting research includes statistical, shallow-learning, and newer deep-learning approaches for dynamic multivariate time-series data. DAQFF extends this direction with an end-to-end hybrid framework targeting local trends and long temporal dependencies.
- Existing air-quality forecasting studies commonly use statistical and shallow-learning methods such as regression, ARIMA, HMM, and artificial neural networks.
- Big-data air-quality forecasting research uses data-driven models to address dynamic and nonlinear time-series characteristics in urban computing.
- Deep-learning work has applied spatial-temporal models, deep recurrent networks, autoencoder pretraining, and unified interpolation-prediction-feature-analysis approaches to air-quality data.
- DAQFF is proposed as an end-to-end hybrid model that uses multivariate time series and automatic feature selection to address local trends and long temporal dependencies.
3 METHODOLOGY
Air quality forecasting is difficult because multivariate pollutant and meteorological series are nonlinear, dynamic, and interdependent across space and time. DAQFF combines multiple 1D-CNNs with Bi-LSTMs to learn local, spatial, and long-term temporal features for PM2.5 prediction.
- Problem formulation: Air quality forecasting targets future PM2.5 concentrations from historical pollutant and related observations across monitoring points.Inputs can include PM2.5, pressure, temperature, wind speed, and other air-quality-related series.
- Problems and motivations: Rapidly changing weather and emissions create nonlinear, dynamic, and interdependent influences that make precise forecasting difficult.Adjacent observations and periodic intervals in pollutant and wind-speed series can be strongly correlated.
- Multiple 1D-CNNs: DAQFF uses multiple one-dimensional CNNs to learn local trends and possible spatial correlations from multiple station time series.The CNN representations are concatenated to capture local features and hidden spatial correlations across stations.
- Bi-LSTM: Bidirectional LSTMs learn spatial-temporal dependencies by processing concatenated features in forward and backward directions.This lets the model use both past and future contexts when forming sequence representations.
- Joint framework: The hybrid framework fuses learned spatial-temporal representations and uses them for final air-quality prediction in an end-to-end architecture.The paper reports good forecasting performance and generalization ability for air-quality prediction tasks.
4 EXPERIMENTS
Experiments on two real-world datasets compare DAQFF with shallow and deep learning baselines for single- and multi-step PM2.5 forecasting. DAQFF maintains the strongest reported performance across datasets, prediction horizons, and varied conditions, while its error depends on lookup size and training epochs.
- Experimental setup: Experiments use two real air-quality datasets and compare DAQFF with classical shallow and baseline deep-learning models.The evaluation includes Beijing PM2.5 data and an Urban Air Quality Dataset.
- Single-step forecasting: DAQFF reduces single-step Beijing errors to RMSE 8.20 and MAE 6.19, while achieving the lowest error on the Urban Air Quality Dataset.The comparison covers ARIMA, SVR variants, RNN, CNN, LSTM, and GRU baselines.
- Parameter analysis: DAQFF reaches its minimum RMSE at a lookup size around 9, whereas larger lookup sizes leave error stable or gradually increase it, potentially from overfitting.Lookup size denotes the historical observation window supplied to the model.
- Parameter analysis: As epochs increase, deep-model RMSE initially decreases; DAQFF remains strongest across epochs, but additional iterations consume resources and can cause overfitting beyond 90 epochs.The epoch analysis uses the Beijing PM2.5 Dataset.
- Single-step forecasting: DAQFF has the lowest single-step prediction error across both datasets and remains best when baseline performance varies substantially between datasets.The paper attributes the advantage to combining CNN-based local-trend extraction with Bi-LSTM long-term dependency learning.
- Multi-step forecasting: DAQFF maintains the best multi-step performance as prediction horizons grow, including short- and long-term forecasts across different datasets and conditions.For six-hour averages, DAQFF reduces Beijing MAE to 27.53 and Urban Air Quality Dataset MAE to 25.01; its advantage also persists at longer horizons.
5 CONCLUSION AND FUTURE WORK
DAQFF is a hybrid CNN and Bi-directional LSTM framework for single- and multi-step PM2.5 forecasting from multivariate air-quality time series. Evaluations on two real-world datasets indicate stronger prediction ability than typical shallow and baseline deep-learning models, while abrupt pollution changes remain a target for future study.
- 5 CONCLUSION AND FUTURE WORK: DAQFF combines one-dimensional CNNs and Bi-directional LSTM to learn local trends, correlation features, and spatial-temporal dependencies in multivariate air-quality time series.The framework performs hierarchical feature representation and multi-scale spatial-temporal dependency fusion in an end-to-end process.
- 5 CONCLUSION AND FUTURE WORK: The study tests DAQFF on two real-world air-quality datasets for both single-step and multi-step forecasting.The experiments compare the proposed model with typical shallow-learning and baseline deep-learning models.
- 5 CONCLUSION AND FUTURE WORK: DAQFF demonstrates better prediction ability than typical shallow-learning and baseline deep-learning models across the evaluated forecasting tasks.The reported evaluation covers both single-step and multi-step forecasting.
- 5 CONCLUSION AND FUTURE WORK: Abrupt changes, also called outlier or anomaly points, in air-pollution time series require further study to improve multi-step forecasting ability.The authors also identify a need to investigate and improve DAQFF under different forecasting conditions.