1. Introduction
The Internet of Things (IoT) is an idea that connects the physical objects to the Internet, which can play a remarkable role and improve the quality of our lives in many different domains [
1,
2]. There are many possibilities and uncertainties in the application scenarios of IoT [
3]. The application of the IoT in the urban area is of particular interest, as it facilitates the appropriate use of the public resources, enhancing the quality of the services provided to the citizens, and minimizing the operational costs of the public administrations, thus realizing the Smart City concept [
4]. The urban IoT may provide a distributed database collected by different sensors to have a complete characterization of the environmental conditions [
2]. Specifically, urban IoT can provide noise monitoring services to measure the noise levels generated at a given time in the places where the service is adopted [
5].
With the unprecedented rate of urbanization as a result of the rapid acceleration of economic and population growth, new problems arose, such as traffic congestion, waste management, pollution, and parking allocation [
6]. Recently, noise pollution has become one of the core urban environmental pollutions and has received increasing attention. Urban noise pollution can cause various consequences, especially severe effects on human health such as hearing damage, affecting sleep and increasing psychological stress [
7,
8]. To solve the problem of noise pollution, the Chinese government issued the Law on the Prevention and Control of Environmental Noise Pollution of the People’s Republic of China [
9] and adopted a series of policy measures on controlling urban noise, rural noise, industrial noise, construction noise, etc. However, due to the variety of noise sources [
10] and the prevalence of noise pollution, the problem of environmental noise pollution in China has not been adequately solved and remained as one of the main urban ecological problems. The Chinese Environmental Noise Pollution Prevention Report showed that in 2017, environmental noise-related complaints received by environmental protection departments at all levels accounted for 42.9% of the total environmental complaints [
11]. According to the latest release of the Beijing–Tianjin–Hebei research report of the environmental resource based on the big data analysis in June 2018, noise pollution cases accounted for 73% of the total number of cases in the Beijing–Tianjin–Hebei regional environmental pollution case [
12]. Therefore, paying attention to urban environmental noise pollution and making an accurate and timely prediction of environmental noise will assist in comprehensively grasping and managing the urban acoustic environment, thus improving residents’ life satisfaction.
Noise prediction has significant theoretical research value and practical significance. In past studies, the authors proposed many methodologies to evaluate and predict the noise in the environment. These prediction methods are mainly classified into three groups: physical propagation models, traditional statistical methods, and machine learning methods.
The first group of noise prediction methods is the physical propagation model, which predicts the ambient noise of another location from the spatial perspective, based on the distance from the sound source to the calculated noise location point and the physical properties of the sound source itself. Since the early 1960s, many developed countries conducted research on airport noise and introduced many airport noise management methods and models [
13,
14], such as the Integrated Noise Model promoted by the Federal Aviation Authority, the Aircraft Noise Contour (ANCON) model of the United Kingdom and the Fluglaerm (FLULA) program developed in Switzerland [
15,
16,
17]. Some studies focused on the spatial propagation of noise to predict environmental noise [
18,
19]. For example, based on the propagation model of noise attenuation, the noise distribution can be drawn [
20]. There are also studies on choosing the location of noise monitoring sites to make the monitoring more effective [
21]. In general, there are few studies focusing on the temporal variation of urban noise using physical propagation models [
18,
19,
20,
21].
The second group of noise prediction methods is the traditional statistical method. Kumar and Jain [
22] proposed an autoregressive integrated moving average (ARIMA) model to predict traffic noise time series, but the amount of data was small, and the monitoring time was only a few hours. It is also concluded that it is better to consider the time series on a longer time scale. Garg et al. [
23] analyzed the long-term noise monitoring data through ARIMA modeling technology and found that the ARIMA method is reliable for time series modeling of traffic noise, but the corresponding parameters should be adjusted according to different situations. Gajardo et al. [
24] performed a Fourier analysis of the traffic noise of the two cities of Cáceres and Talca, and discovered that regardless of the measurement environment of the city and the corresponding traffic flow, larger periodic components and amplitude values are similar in different samples, which indicated that urban traffic noise has an inherent law that can be predicted. There is also a study that developed models based on the regression model to prognosticate the level of noise in closed industrial spaces through the dominant frequency cutoff point [
25].
The third group of noise prediction methods is the machine learning method. Rahmani et al. [
26] introduced the genetic algorithm into the traffic noise prediction and proposed two prediction models, and the simulation on the actual data achieved an accuracy of ±1%. Wang et al. [
27] occupied a back propagation neural network to simulate traffic noise. Iannace et al. [
28] used a random forest regression method to predict wind turbine noise. Torija and Ruiz [
29] used the feature extraction and machine learning methods to predict the environmental noise and got an excellent fitting effect. However, the number of input variables was 32, the distinction is subtle, and it is challenging to organize the data.
The prediction of environmental noise heavily depends on historical and real-time noise monitoring data. The research by Frank et al. [
30] shows that combining the rules or patterns mined in the monitoring data with the acoustic theoretical calculation model can effectively improve the prediction accuracy of noise. In terms of reflecting regional noise levels, sometimes sampling strategies are put forward based on considerations of saving resources and improving data acquisition efficiency [
31]. Giovanni et al. [
32] found that considering saving the time cost and the accuracy of the results, a non-continuous 5–7 days observation of noise is reasonable for long-term noise prediction.
Forecasting the temporal variation of noise can offer a scientific basis for urban noise control. In recent years, with the widespread utilization of sound level meters and the development of various sensor network technologies, environmental noise data has an exploding expansion. Although there have been previous studies on noise measurement, prediction, and control [
19,
22,
23,
25,
33,
34,
35,
36], most of the research data are relatively diminutive. This gave the inspiration to have a second thought about the environmental noise prediction problem, that is, whether there are more optimized noise prediction models and methods when handling an abundant amount of noise data. Therefore, predicting noise in the time dimension requires a more efficient approach. However, at a fine time interval, few studies have focused on and predicted the variation of noise within a given day so far.
Deep learning has developed rapidly and has been successfully applied in many specialties lately [
37]. It utilizes multiple-layer architectures or deep architectures to extract inherent features in data from the lowest level to the highest level, and they can discover huge amounts of structure in the data [
38]. Deep learning is derived from the study of artificial neural networks. The most common neural network models include multilayer perceptron (MLP), convolutional neural networks (CNN), recurrent neural networks (RNN), etc. [
39]. For time series, RNN is often employed to characterize the association of hidden states and capture the data characteristics of the entire sequence. Nevertheless, simple RNN has long-term dependence problems and cannot effectively utilize long-interval historical information. Therefore, long short-term memory (LSTM) network has emerged to unravel the problem of gradient disappearance, which has been used for stock price forecasting [
40], air quality forecasting, sea surface temperature forecasting [
41], flight passenger number forecasting, and speech recognition [
42]. The results illustrated that the model had achieved excellent performance.
In this study, we deployed an IoT-based noise monitoring system to acquire the urban environmental noise, and proposed an LSTM network to predict the noise at different time intervals. The performance of the model was compared with three classic predictive models—random walk (RW), stacked autoencoder (SAE), and support vector machine (SVM) on the same dataset. This study also explored the impact of monitoring point location on prediction results and policy recommendations for environmental noise management.
5. Conclusions
Conclusively, in the context of the rapid development of the environmental IoT, this study proposed a general method for predicting the timely environmental noise via the application of the LSTM network in a large data volume scenario and compared it with the performance of the classical model on the same data set, verifying the feasibility and effectiveness of the LSTM neural network in predicting environmental noise. However, due to the limitation of the duration of data collection time, this prediction has no way to verify the predictive ability of the LSTM network on the daily/monthly noise level. Nonetheless, as the time interval increases, the average noise level is more stable, and the randomness declines. Under the condition that the training samples are sufficient, the performance of the prediction model should be better. Furthermore, from this study, it is believed that the LSTM network can be applied simply to other noise data sets on predicting environmental noise, but different data sets may need to be re-adjusted. The shortcoming of this study is that the LSTM network structure used is relatively simple, and in the future, a more in-depth, broader and more powerful optimized LSTM model can be designed to improve accuracy. At the same time, the LSTM does not provide insight into the physical meaning of their parameters [
62], in addition to time, more variables should be considered.
China’s past environmental noise monitoring method is mostly based on the statistical average. The monitoring data can only represent the overall average level of the specific time and place, which lacks understanding and application of environmental noise characteristics and acoustic laws, and in addition, lacks analysis of causes and effects. Establishing the connection between environmental noise and position, time, traffic, population, and other factors, replacing part of the monitoring work with prediction is the development direction of environmental noise evaluation. The monitoring and evaluation methods should be optimized from this perspective to make the monitoring results more in-depth and representative.