Abstract
With the rapid development of Internet of Things (IoT) technologies, smart city public information platforms are collecting increasing volumes of urban data. Leveraging this data to analyze the "pulsation" of city development is crucial for addressing urban challenges. This paper explores data sources for environmental pulsation analysis within the Sino-Singapore Tianjin Eco-city (SSTEC). By analyzing time-series features, we implemented a wavelet-enhanced ARMA model to predict air pollutant concentrations and water quality. Results indicate that this hybrid approach offers superior prediction accuracy, providing valuable guidance for urban environmental management and other smart city applications.
Introduction: The Challenge of Urban Pulsation
As smart cities deploy extensive IoT sensor networks, they generate a massive influx of spatial-temporal data. However, urban planners are often "data rich but information poor." This research targets the concept of urban environmental pulsation—the rhythmic, multi-scale patterns of change in a city's environment, such as daily traffic cycles versus seasonal climate shifts.
Traditional time-series models like ARMA (Auto-Regressive Moving Average) excel at capturing global linear trends but often fail to account for short-term, unexpected fluctuations or non-linear noise. Conversely, pure machine learning models can struggle with the underlying stochastic nature of environmental data. The challenge lies in developing a model that simultaneously handles broad trends and fine-grained noise.
Methodology: A Hybrid Forecasting Framework
The proposed solution is a Wavelet Decomposition enhanced ARMA model. This approach employs a "divide and conquer" strategy to process complex environmental signals.
The process involves four key stages:
- Wavelet Multi-scale Decomposition: Using the Mallat algorithm (specifically the
db3wavelet), the raw, noisy environmental signal is decomposed into multiple layers. This produces an approximation sequence (low-frequency global trend) and several detail sequences (high-frequency fluctuations and noise). - Stationarity & Linearity Testing: The Augmented Dickey-Fuller (ADF) test ensures data stationarity, while the McLeod-Li test confirms non-linearity, justifying the need for a hybrid approach over standard linear regression.
- Individual ARMA Modeling: A separate ARMA model is fitted to each decomposed sequence, allowing the system to capture the unique characteristics of both the trend and the noise independently.
- Reconstruction: Predicted values from all individual models are recombined using the Mallat reconstruction algorithm to generate the final forecast.
Data and Experimental Setup
The study utilizes real-world data from the Sino-Singapore Tianjin Eco-city (SSTEC):
- Air Quality: Data from the Luzhuang Garden monitoring station (2009–2015), covering $SO_2$, $NO_2$, $CO$, $O_3$, $PM_{10}$, and $PM_{2.5}$.
- Water Quality: Data from 13 monitoring sites (2012–2015), including pH, suspended solids, COD, TOC, nitrogen ammonia, and phosphorus.
Key Insights: The "Prism" Effect
The core innovation lies in treating Wavelet Decomposition like a prism. Just as a prism splits white light into its constituent colors, wavelets split a messy environmental signal into different "frequencies" of city life. Modeling these frequencies separately is far more effective than attempting to model the aggregate signal. Interestingly, while the raw data is non-linear, the decomposed sequences are more "well-behaved," allowing the linear ARMA model to perform exceptionally well.
Results and Performance
The hybrid model demonstrated robust performance across various pollutants. In head-to-head tests on $PM_{2.5}$ data, the proposed method achieved an RMSE of 0.068, significantly outperforming standard ARMA (0.076), ANN (0.088), and SVR (0.097).
The Mean Relative Error (MRE) for training data remained consistently below 25%, indicating a high-quality fit to the historical "pulsation" of the city. The model was particularly effective for nitrogen ammonia and phosphorus in water quality tests, likely due to the higher stationarity of those specific chemical sequences.
Conclusion and Future Directions
While the wavelet-enhanced ARMA model shows great promise, the authors noted slight over-fitting, as test errors were higher than training errors. Additionally, monthly predictions were less accurate than daily ones due to the "average effect," which smooths out the fine-grained pulsation information. Future work will explore applying this framework to other smart city domains, such as transportation and energy management.
Summary updated