Keywords

1 Introduction

Climate change is a modern topic nowadays. Air pollution is one of the most important environmental problems on the globe, and causes many types of allergies, respiratory illnesses, cardiovascular diseases, acute bronchitis diseases, etc. [1, 2]. Particulate matter (PM) is an air pollutant with high impact on humans because short-term and long-term exposure to high concentrations may produce severe health effects and premature mortality [3, 4].

Short-term forecasting of PM2.5 air pollution trends can use different methods: deterministic, statistical, neural, hybrid (e.g. neuro-fuzzy) etc. The statistical models include linear regression, ARIMA, principal components analysis, etc., and have been used for their forecasting skills [5, 6]. The forecasted results generated using these linear statistical models are in general not satisfactory. An alternative is the use of computational intelligence approaches, such as artificial intelligence-based models [5, 7]. Artificial neural networks [8] and adaptive neuro-fuzzy inference systems (ANFIS) have been successfully applied in air pollution forecasting domain [9–11]. The chosen of an efficient forecasting method is done by experiment, depending on the available time series databases with measurements of PM2.5 concentration, meteorological parameters, other air pollutants concentration that influence PM2.5. Depending on the correlation degree with PM2.5, a part of these parameters can be considered as inputs in the PM2.5 forecasting model. We are applying such a model under the ROKIDAIR research project (http://www.rokidair.ro) whose goal is to provide an intelligent tool (ROKIDAIR DSS) for early warning/alerting of PM2.5 air pollution episodes in urban areas (in two pilot cities from Romania, Ploiesti and Targoviste), in order to reduce the potential negative effects of air pollution on children health. Within this project we are developing a model based on artificial intelligence, named ROKIDAIR IA which has two main components: a short-term PM2.5 forecasting component and an intelligent decision support component, based on knowledge. In this paper we focus on short-term PM2.5 forecasting modeling based on ANN.

2 The Artificial Neural Network Approach for Short-Term PM2.5 Forecasting

Artificial neural networks are universal approximators that can learn complex mapping between the input and the output data [12]. An ANN is composed by a set of artificial neurons which are connected according to a topology. Each connection between two neurons has a weight (a numerical value in the interval [0, 1]) showing the degree of that connection which is derived during the ANN training stage. The number of input neurons is given by the input parameters of the forecasting problem, the output neurons are the PM2.5 forecasted values in the time window t + k (named also, forecast horizon), while the number of hidden neurons is derived by experiment during training. Some of the ANNs types most used to solve forecasting problems are feed forward artificial neural networks [13], recurrent ANNs [14] and radial basis ANNs [12]. Some recent research results reported in the literature confirmed the good performance of the neural predictors used to detect the air pollution evolution [15–17].

Figure 1 shows an example of a feed forward ANN for PM2.5 forecasting. The model uses past measurements of PM2.5 concentration and other atmospheric parameters. The ANN has an input layer, an output layer and one or more hidden layers. Usually, one hidden layer is enough to capture the evolution of the forecasted parameter according to the data sets available for ANN training. Feed forward ANNs are trained with a backpropagation algorithm which can be improved by choosing the right learning parameters, adjusted during training. The generation of an ANN model must follow three steps: (1) ANN training with a training algorithm on a training data set; (2) ANN validation on a training data set; (3) ANN testing on a testing data set.

Fig. 1.
figure 1

Example of a feed forward ANN for PM2.5 forecasting

Full size image

The PM2.5 ANN forecasting model is derived by training the ANN on a training set selected from the data sets that are available for the urban area that is studied. After training the ANN model is validated and tested on specific data sets. A recent comparison between some ANN models applied to PM2.5 prediction is described in [18]. The main advantage of an ANN forecasting model is given by its capability to capture with good accuracy the forecasting function when enough large data sets are used. Our proposed approach for PM2.5 short-term forecasting is based on the MBP - PM 2.5 forecasting protocol, developed under the ROKIDAIR project.

3 The PM2.5 Forecasting Model Development Protocol

We have developed a protocol, MBP - PM 2.5 forecasting, for building the PM2.5 forecasting model under the ROKIDAIR project. The main purpose of the protocol is to facilitate the systematic construction of the short-term PM2.5 forecasting model that will be used by the ROKIDAIR Decision Support System in order to provide decisions under the form of warning/alerting messages regarding the potential negative effects on children health of the PM2.5 air pollution episodes. The MBP - PM 2.5 forecasting protocol defines the steps of PM2.5 forecasting model design. The air pollution forecasting module determines the evolution for short term PM2.5 concentration.

Figure 2 presents the logic diagram of the MBP - PM 2.5 forecasting protocol (with 4 main steps) for the short-term PM2.5 forecasting module of the ROKIDAIR Decision Support System.

Fig. 2.
figure 2

Logic diagram of the MBP-PM 2.5 forecasting protocol (steps I, II, III, IV)

Full size image

In the first step are set the PM2.5 forecasting model requirements (as e.g. past measurements window time, forecasting horizon, input/output parameters, forecasting accuracy). In step II it is checked if the database with PM2.5 concentration measurements and other PM2.5 related atmospheric parameters measurements (DB-PM2.5) has enough data. If there are not enough data, it is started a process of data collection (usually, for the analyzed urban area or a similar PM2.5 air polluted urban area). When enough data are stored in the database, step III is performed with a data analysis sub-step (III.1), followed by the setting of the forecasting method (III.2) and the design of the PM2.5 forecasting model (III.3) according to the methodology of selecting the best solution. After step III, the short-term PM2.5 forecasting model is generated. If the model is not validated, another forecasting method is chosen in step III.2. If the model is validated than it is adopted by the ROKIDAIR system. The model validation is performed according to the desired forecasting performance which is measured with some indicators: mean absolute error (MAE), index of agreement (IA), root mean square error (RMSE), and coefficient of determination (R2).

As we are focusing on the artificial neural network based forecasting method, we present the main steps of the methodology proposed for feed forward and radial basis ANN model selection which were integrated in the ROKIDAIR MBP – PM 2.5 forecasting protocol.

MBP – PM 2.5 forecasting protocol – ANN Selection Methodology

  • Step 1. Time series data processing (i.e. DB-PM2.5) – in order to be used by the PM2.5 ANN forecasting method;

  • Step 2. Select the most relevant atmospheric input parameters to short-term PM2.5 forecasting (e.g. by using principle component analysis);

  • Step 3. Select the training, validation and testing data sets for the ANN model;

  • Step 4. Set the ANN architecture (e.g. input nodes, output nodes, hidden nodes, radial function, cluster seed, number of clusters etc.);

  • Step 5. Adjust the training parameters according to the training algorithm;

  • Step 6. ANN training, validation, testing using the training data set chosen in step 3;

  • Step 7. Analyze the performances of the designed ANN model (i.e. RMSE, IA, R2);

  • Step 8. Select the best ANN model for real time short-term PM2.5 forecasting.

A good PM2.5 forecasting model should have a smaller error (RMSE, MAE), a coefficient of determination and an index of agreement close to 1. In order to keep the PM2.5 short-term forecasting model as simple as possible for an efficient real time PM2.5 forecasting, a minimum number of the atmospheric parameters (e.g. temperature and relative humidity) most relevant to PM2.5 concentration evolution are chosen.

4 Experimental Results

The data sets used in this study come from an air quality monitoring station from an urban area of Ploiesti, Romania, and each data set contains approximately 4200 samples for PM2.5 concentrations and temperature. From all meteorological parameters the temperature is correlated with PM2.5 evolution. The data from Ploiesti monitoring station referring to PM2.5 concentrations has the maximum of 36.45 μg/m3, and a minimum of 0.19 μg/m3. In the same time, the temperature data set has the maximum of 37.24 °C, and the minimum of −0.2 °C.

The proposed forecasting models use normalized data for both PM2.5 concentrations and temperature. The data were randomly divided with the following percentages: 70 % for training, 15 % for validation and 15 % for testing. We propose two types of forecasting models in this study, based on ANNs. One model has as inputs the four previous PM2.5 hourly concentrations (Fig. 3a) and the other has one more input than the first one, namely the current hourly temperature (Fig. 3b). The output of the models is the same in both cases - short term forecasted value for the next hour PM2.5 concentration.

Fig. 3.
figure 3

ANN models

Full size image

The structure of the proposed neural network contains four neurons in the input layer, one hidden layer and one neuron in the output layer. In the study there were used two types of neural networks, namely feed forward backpropagation (FFwd) and layer recurrent (LRec). As training algorithm the preferred method is Levenberg-Marquardt, and for the adaptive learning functions there are studied the gradient descent with momentum weight and bias (learngdm) and gradient descent weight and bias (learngd). The simulations were performed modifying also the number of neurons in the hidden layer.

The training and validation errors have values around 0.001 and 0.0007 respectively. The accuracy of the models can be evaluated based on the comparison between the actual value and forecasted value of PM2.5 concentration, with mean error and standard deviation criteria. The performances of the designed ANN models are compared using statistical indices such as RMSE, IA, R2, and R.

The two models are compared using statistical criteria and a selection of the results are presented in Tables 1 and 2, the best configuration for each ANN model being highlighted.

Table 1. Statistical indices for ANN model 1
Full size table
Table 2. Statistical indices for ANN model 2
Full size table

For the first model using only PM concentrations as inputs the best results are obtained in the case of layer recurrent structure with 5 neurons in the hidden layer and the learngdm adaptation learning function. In this case the root mean squared error have the smallest value, and IA, R2 and R indices have the biggest values.

The second model with temperature as additional input has the best results (comparing the same statistical indices) in the case of feed forward structure with 6 neurons in the hidden layer and the learngd adaptation learning function.

The best results from the two models showed that no significant enhancement has been produced when current hourly temperature is included as additional input variable to the second ANN model. The best structure between the two is the one from the first model with PM concentrations as inputs (4 × 5 × 1 – Learngdm – Layer Recurrent) with: RMSE = 1.0908 μg/m3, IA = 0.9905, R2 = 0.9634 and R = 0.9815.

Figure 4 presents a partial view of the comparison between testing and forecasted data for the best ANN structure.

Fig. 4.
figure 4

Comparison between testing and forecasted data

Full size image

5 Conclusions

The paper presented two ANN models proposed for real time PM2.5 short-term forecasting in the case of a polluted town in Romania. In order to select the best ANN forecasting model, we have designed a model based PM2.5 forecasting protocol, named MBP – PM 2.5 forecasting protocol, which is integrated in the ROKIDAIR MBP protocol for the development of the ROKIDAIR DSS. The first proposed model uses as inputs only hourly PM concentrations and the second one uses an additional input the current hourly temperature. The conclusions are that the accuracy of both ANN models are almost the same, so both models can be considered appropriate approaches to real time short term forecast. As future work we propose to include other meteorological variables into the model, use additional hybrid modelling techniques such as FIR with genetic algorithm, or expand the forecasting window to next day.