Real-time Prediction for Fine-grained Air Quality Monitoring System with Asynchronous Sensing

Zixuan Bai, Zhiwen Hu, Kaigui Bian, Lingyang SongICASSP 2019 · OpenReview

TL;DR

This paper tackles the challenge of predicting fine-grained air quality in urban areas using data from a network of IoT sensors that operate asynchronously to conserve energy. They introduce a novel "correlation graph" model to capture the complex spatial, temporal, and meteorological relationships within this sparse, asynchronous data, leading to a more accurate real-time prediction system.

What Problem Does This Paper Address?

Air pollution is a major global health concern, and accurate, real-time monitoring and prediction of air quality are crucial. Traditional monitoring stations offer coarse-grained data with significant latency. While deploying dense networks of low-cost IoT sensors can provide fine-grained measurements, these sensors are often battery-powered and must operate asynchronously to prolong their lifespan. This asynchronous operation results in sparse and irregular data, which existing prediction methods like LSTMs, MLPs, and Inverse Distance Weighting (IDW) struggle to handle effectively. These methods often assume synchronous data or fail to adequately model the intricate spatial-temporal and meteorological dependencies inherent in urban air quality.

Methods

The core of the proposed solution lies in a novel Correlation Graph (CG) model. This CG represents the air quality monitoring system at any given time slot t. It comprises multiple subgraphs, each representing a specific time window (past, present, and future). Each node in the graph corresponds to a Point of Interest (POI) and is either labeled (sensor data available) or unlabeled (data needs prediction).

The CG captures relationships through:

  • Weighted Feature Vectors: Each node's features (spatial coordinates, temporal information, and meteorological conditions) are weighted using a learned vector β to reflect their importance.
  • Node Similarity: An adjusted cosine similarity metric is used to quantify the relationship between any two nodes.
  • Weighted Edges: Edges between nodes are weighted based on their spatial-temporal proximity and similarity, with a hyperbolic tangent function used to magnify differences in similarity.

The prediction process involves two main procedures:

  1. Preparation Procedure: This step generates a "pre-estimation" vector Yt for all nodes. Labeled nodes are assigned their measured values. Unlabeled nodes not in the newly added subgraph inherit their previous time slot's estimated value. For unlabeled nodes in the newly added subgraph, an LSTM network is used to predict their coarse-grained estimation based on historical meteorological data and coarse-grained AQI estimates.
  2. Estimation Procedure: This procedure refines the pre-estimated values into precise predictions Ft by minimizing a loss function that balances smoothness (nearby nodes should have similar values) and reliability (predictions should be close to pre-estimations). This minimization problem is solved using a combination of closed-form derivation for the prediction values Ft (given β) and a genetic algorithm to optimize the feature weight vector β based on long-term data.

Datasets

The paper utilizes data collected from their own Internet-of-Things-based sensing system in Peking University, deployed since February 2018. This dataset includes AQI (specifically PM2.5 values), sensor locations, timestamps, and weather conditions obtained from the China Meteorological Administration.

Key Insights and Ideas

  • Modeling Asynchronous Data: The CG approach is a clever way to represent and leverage sparse, asynchronous sensor data by explicitly modeling temporal relationships across multiple time slots.
  • Spatial-Temporal-Meteorological Fusion: The weighted feature vectors and edge weights effectively integrate diverse data sources (location, time, weather, AQI) into a unified graph structure.
  • Genetic Algorithm for Feature Weighting: Using a genetic algorithm to learn the optimal feature weights (β) is a robust way to adapt the model to the specific characteristics of the air quality data and its influencing factors.
  • Balancing Smoothness and Reliability: The loss function design is crucial for ensuring that predictions are both locally consistent (smooth) and grounded in the available measurements (reliable).

Key Results

The proposed CG model demonstrates significant performance gains over existing methods like MLP and Inverse Distance Weighting (IDW) for fine-grained air quality prediction with asynchronous sensing. In simulations, the CG model achieved an average relative error of 5% when using an appropriate number of history (Th=5) and future (Tf=3) subgraphs and 30 sensors. This is a substantial improvement compared to MLP, which achieved around 9% error under similar conditions. The paper also highlights that the number of history and future subgraphs (Th and Tf) needs to be carefully tuned for optimal performance, as too few or too many can degrade accuracy.

Limitations and Future Work

  • Parameter Tuning: The performance of the CG model is sensitive to the choice of parameters, particularly the number of history and future subgraphs (Th and Tf). The paper suggests that finding the optimal balance is key.
  • Computational Cost: While the genetic algorithm is effective for optimizing β, it can be computationally intensive.
  • Generalizability: The study is conducted on data from Peking University. Further validation on datasets from different urban environments with varying geographical and meteorological conditions would be beneficial.
  • Future Work: The authors imply that further exploration of the genetic algorithm's efficiency and the model's adaptability to different urban settings would be valuable.

Summary updated

Abstract

Due to the significant air pollution problem, monitoring and prediction for air quality have become increasingly necessary. To provide real-time fine-grained air quality monitoring and prediction in urban areas, we have established our own Internet-of-Things-based sensing system in Peking University. Due to the energy constraint of the sensors, it is preferred that the sensors wake up alternatively in an asynchronous pattern, which leads to a sparse sensing dataset. In this paper, we propose a novel approach to predict the real-time fine-grained air quality based on asynchronous sensing. The sparse dataset and the spatial-temporal-meteorological relations are modeled into the correlation graph, in which way the prediction procedures are carefully designed. The advantage of the proposed solution over existing ones is evaluated over the dataset collected by our air quality monitoring system.
BibTeX
@inproceedings{bai2019real,
  author = {Zixuan Bai and Zhiwen Hu and Kaigui Bian and Lingyang Song},
  title = {{Real-time Prediction for Fine-grained Air Quality Monitoring System with Asynchronous Sensing}},
  booktitle = {IEEE International Conference on Acoustics, Speech and Signal Processing},
  year = {2019},
  url = {https://openreview.net/forum?id=j1T3KKS8BK},
}
Figure