KR20230123574A - Transfer learning driven sequential forecasting and ventilation control of PM2.5 associated health risk levels in underground public facilities - Google Patents

Abstract

The present invention relates to a fine dust prediction system and prediction method using artificial intelligence transfer learning, and a ventilation control system and method therethrough, and more particularly, a prediction model generator for building a prediction model based on transfer learning; And a fine dust prediction unit for predicting the concentration of fine dust after a specific time based on the prediction model built through the prediction model generation unit; including, using the transfer learning-based ResNet prediction model as a prediction model, data in another space About a fine dust prediction system using artificial intelligence transfer learning, characterized by predicting the PM _2.5 concentration after a specific time in the prediction target space by learning the prediction system even when the learning data of the prediction target space is insufficient by using will be.

Description

Fine dust prediction and ventilation control system and method using artificial intelligence transfer learning {Transfer learning driven sequential forecasting and ventilation control of PM2.5 associated health risk levels in underground public facilities}

The present invention relates to a fine dust prediction system and prediction method using artificial intelligence transfer learning, and a ventilation control system and method through the same.

Fine particulate matter (PM) has been a constant topic in recent years due to its associated health impacts, and reducing its impacts is a global challenge. In response, the United Nations (UN) has emphasized practical measures to reduce deaths and diseases caused by fine dust as part of the Sustainable Development Goals agenda. Among them, PMs with aerodynamic diameters smaller than 2.5 μm (PM _2.5 ) and 10 μm (PM ₁₀ ) have attracted particular attention because they can cause lung and cardiovascular diseases when inhaled. Exposure to PM is closely related to the indoor microenvironment in which people spend more than 80% of their time. Therefore, it is important to analyze, monitor and mitigate the health effects of PM in these indoor environments.

The toxicity of PM penetrates the inside of the lungs and directly interacts with cellular structures, causing lung inflammation. Depending on its size, it has been established that PM _2.5 can accumulate toxic particles deep in the lungs, erode alveolar walls and impair lung function. Moreover, long-term exposure to PM _2.5 is associated with increased cardiovascular mortality and morbidity. In this context, the American Cancer Society concluded that for every 10 μg/m ³ increase in PM _2.5 concentration, there was a 4%, 6%, and 8% increase in total adult mortality, cardiovascular mortality, and lung cancer, respectively. .

In recent years, metro systems have been rapidly developed in large cities around the world to alleviate high traffic congestion and vehicle exhaust pollution. Millions of people use the metro system as their primary commute because of its convenience. Despite these benefits, the effect of the indoor microenvironment on passenger health in these underground metro systems has generated considerable interest from government and academia. Modern subway systems have relatively closed and complex characteristics with limited natural ventilation conditions and emission source characteristics of PM. Particulate matter also represents a complex of many components that are influenced by carbon mass, carbon element, metal particles, train wheel track wear, outdoor contaminant penetration, train frequency and passenger density.

The Ministry of Environment (MOE) classified subway stations as indoor hazardous facilities. Accordingly, the Ministry of Environment prepared measures for air quality monitoring, such as implementing the Indoor Exposure Act, adding air quality monitoring stations, and installing indoor sensors to monitor abnormal situations. Contaminant concentration levels are graded from "good" to "very bad" to provide intuitive information. Each rating represents an associated health risk for different target groups. Additionally, the US Environmental Protection Agency (EPA) has proposed the Comprehensive Indoor Air Quality Index (CIAI) to determine the level of public health risk associated with indoor air quality (IAQ) levels.

To ensure sustainable IAQ levels, forecasting air pollution information is essential to developing early warning systems for effective response. For example, this information helps to obtain pre-inverter frequencies to regulate subway ventilation control systems as time is needed to dissipate current PM levels. In previous studies, numerous techniques have been used for the analysis, prediction and control of PM, including statistical and mathematical models. However, considering the complex and non-linear relationship between PM and other measurable variables in the IAQ system, existing regression methods cannot capture the dynamic PM distribution and have poor predictive performance. In contrast, artificial neural networks (ANNs) incorporating nonlinear activation functions are better able to capture these dynamic behaviors.

Recurrent neural networks (RNNs), such as long short-term memory (LSTM) and gated recurrent unit (GRU) networks, have been widely applied for PM prediction. Nonetheless, it is difficult to train deep RNNs to learn unique features from data because of the performance degradation problem caused by vanishing gradients in deep networks. However, recent advances in Convolutional Neural Networks (CNNs), especially Residual Neural Networks (ResNets), have enabled deep model training through mapping recognition. RNNs are mostly implemented to capture the periodicity of time series data, but ResNet shows excellent performance for time series forecasting applications such as PM _2.5 .

In practice, time series data change depending on weather conditions and seasonality, resulting in variability between old and new data, making it a forecasting approach that is only applicable under certain conditions. In the case of other subways in the metropolitan area, relatively few data are available due to regular monitoring, difficult monitoring environments, or expensive measuring equipment. This presents new limitations in predicting future air quality information when data are insufficient, and this problem needs to be overcome.

The aforementioned problem of subway PM prediction is closely related to a branch of machine learning known as transfer learning. This framework enables the use and transfer of knowledge about work from one domain to other related domains. Previously, transfer learning was applied to solve the prediction problem of buildings. For example, we used a transfer learning framework for energy prediction of buildings. They conducted experiments with different levels of data availability and concluded that a transfer learning-based methodology could reduce prediction errors by up to 78%. Grubinger et al. (2017) used generalized online transfer learning for temperature prediction of residential buildings. They concluded that learning from multiple similar sources can improve the predictive performance and generalization ability of the target domain. Also, regular RNNs are vulnerable to the lack of data problem in the training phase. Therefore, it shows a lower suitability for discriminating the temporal order of the PM distribution compared to the pre-trained transfer learning-based transformation.

Existing artificial intelligence-based prediction systems rely heavily on accumulated sensor data, and have the disadvantage that performance degradation may occur when a new monitoring environment is built and sensors are installed.

Korean Registered Patent No. 10-2196091 Republic of Korea Patent Publication 10-2021-0108826 Republic of Korea Patent Publication 10-2021-0086786 Republic of Korea Patent Publication 10-2019-010693

Therefore, the present invention has been devised to solve the above conventional problems, and according to an embodiment of the present invention, transfer learning (transfer learning, TL) based residual neural network (ResNet) based on artificial intelligence technology Its purpose is to provide a PM _2.5 prediction system for underground stations.

According to an embodiment of the present invention, the PM _2.5 concentration after 1 hour is predicted using transfer learning-based ResNet, and the measured PM _2.5 data of the underground station are available at 20%, 40%, 60%, and 80%, respectively. It was confirmed that the performance was improved compared to the existing prediction system through the evaluation of the prediction performance using TL-ResNet for the scenario of the time.

And according to the embodiment of the present invention, it is possible to improve the prediction performance by more than 40% by verifying using data in domestic underground stations, and when applied to the ventilation system, the indoor PM _2.5 concentration is improved by 29% using the predicted data. The purpose is to provide a fine dust prediction, ventilation control system and method using artificial intelligence transfer learning that can be performed.

In addition, according to the embodiment of the present invention, it is possible to predict and improve indoor air quality by applying it to domestic and foreign underground station ventilation process systems, and it can be expected to be applied to various ventilation control fields. For commercialization, domestic and foreign underground stations and environmental systems The purpose is to provide a fine dust prediction, ventilation control system and method using artificial intelligence transfer learning, which can be used for indoor environment improvement in companies, etc.

On the other hand, the technical problems to be achieved in the present invention are not limited to the above-mentioned technical problems, and other technical problems that are not mentioned will become clear to those skilled in the art from the description below. You will be able to understand.

An object of the present invention is a fine dust prediction system, comprising: a prediction model generator for building a prediction model based on transfer learning; And a fine dust prediction unit for predicting the concentration of fine dust after a specific time based on the prediction model built through the prediction model generation unit; can be achieved as a fine dust prediction system using artificial intelligence transfer learning, characterized in that it includes .

And the prediction model may be characterized in that it is a ResNet prediction model based on transfer learning.

In addition, the specific time may be 40 to 100 minutes, and the fine dust data may be PM _2.5 data.

And, the ResNet prediction model may be characterized by applying source data within a space other than the corresponding space to be predicted.

In addition, using the ResNet prediction model, it is possible to predict the PM _2.5 concentration after a specific time in the prediction target space by learning the prediction system even when the learning data of the target space to be predicted is insufficient by using the source data of other spaces. can be characterized.

Further, the ResNet structure may include a convolution block, an identity block, a fully connected layer, and an output layer.

In addition, the ResNet prediction model may be derived through pre-training on source data, and may be characterized in that it is fine-tuned for data available in the target space.

In addition, the ResNet prediction model may be characterized in that the initial layer of the source data is fixed, the PM distribution is stored in sufficient data, and the loss during fine-tuning is minimized and applied to the target space.

In addition, the ResNet prediction model fixes the feature extractor parameters of the pre-trained model after pre-training the source data, and the fixed layer is in a state in which the weight of the layer is not changed when the predictive model is assigned to perform another task. , and the remaining classifier layers are fine-tuned for the target data in the target space.

And the fixed layer may be characterized in that 35 ~ 45%.

According to an embodiment of the present invention, it is possible to provide a PM _2.5 prediction system of an underground station based on a residual neural network (ResNet) based on transfer learning (TL) among artificial intelligence technologies.

According to an embodiment of the present invention, the PM _2.5 concentration after 1 hour is predicted using transfer learning-based ResNet, and the measured PM _2.5 data of the underground station are available at 20%, 40%, 60%, and 80%, respectively. It was confirmed that the performance was improved compared to the existing prediction system through the evaluation of the prediction performance using TL-ResNet for the scenario of the time.

In addition, according to the fine dust prediction and ventilation control system and method using artificial intelligence transfer learning according to an embodiment of the present invention, the prediction performance can be improved by more than 40% by verifying using data in domestic underground stations, and the ventilation system When applied, it has the effect of improving the indoor PM _2.5 concentration by 29% using the predicted data.

In addition, according to the fine dust prediction and ventilation control system and method using artificial intelligence transfer learning according to an embodiment of the present invention, indoor air quality can be predicted and improved by applying to domestic and foreign underground station ventilation process systems, and various ventilation controls It can be expected to be applied to the field, and for commercialization, it has the advantage of being used for indoor environment improvement at domestic and foreign underground stations and environmental system companies.

On the other hand, the effects obtainable in the present invention are not limited to the effects mentioned above, and other effects not mentioned will be clearly understood by those skilled in the art from the description below. You will be able to.

The following drawings attached to this specification illustrate preferred embodiments of the present invention, and together with the detailed description of the invention serve to further understand the technical idea of the present invention, the present invention is limited only to those described in the drawings. and should not be interpreted.
1 is a block diagram of a fine dust prediction system using artificial intelligence transfer learning according to an embodiment of the present invention;
Figure 2 is a schematic diagram of the basic principle of transfer learning
3 is a building block matching of Deep ResNet (deep residual network);
4a shows changes in air quality data measured at E4 subway station;
Figure 4b is the location of the target subway station on line 2 of the SMS.
Figure 5a is a C-subway station, Figure 5b is an E3-subway station, Figure 5c is an E4-subway station, and Figure 5d is an outdoor/indoor PM10 and PM2.5 measurement distribution graph of a D-subway station,
Figure 6 is a statistical correlation analysis of measurement variables and platform PM2.5 for function selection,
Figure 7 is a framework of ResNet: (a) convolution and ID block with activation function, (b) the overall structure of ResNet,
8 is a graphical representation of a training process for a transfer learning framework in accordance with an embodiment of the present invention;
9 is a schematic diagram of a methodology performed for sequence prediction of PM2.5 CIAI at a subway platform according to an embodiment of the present invention;
10 is a schematic representation of a transfer learning-based framework proposed for sequential prediction and control of the platform PM2.5 CIAI level according to an embodiment of the present invention;
11 is a sequence prediction of platform PM2.5 CIAI by moving window in E3-subway according to an embodiment of the present invention: (a) MLR, (b) S-LSTM2, (c) S-GRU2, (d) S -LSTM10, (e) S-GRU10 and (f) ResNet;
12 shows multi-sequence prediction performance of standalone ResNet (left) and TL-ResNet (right) for platform PM2.5 CIAI according to an embodiment of the present invention: (a) E4 subway, (b) C subway, (c) D subway station,
13 is a graph of RMSE scores and performance improvement of standalone ResNet and TL-ResNet on E4, C and D-subway using different data availability scenarios;
Figure 14 shows ventilation response and CIAI level analysis of subway platform for 40% available IAQ data: (a) predicted data, (b) performance of standalone ResNet, (c) performance of TL-ResNet.

The above objects, other objects, features and advantages of the present invention will be easily understood through the following preferred embodiments in conjunction with the accompanying drawings. However, the present invention is not limited to the embodiments described herein and may be embodied in other forms. Rather, the embodiments introduced herein are provided so that the disclosed content will be thorough and complete and the spirit of the present invention will be sufficiently conveyed to those skilled in the art.

In this specification, when an element is referred to as being on another element, it means that it may be directly formed on the other element or a third element may be interposed therebetween. Also, in the drawings, the thickness of components is exaggerated for effective description of technical content.

Embodiments described in this specification will be described with reference to cross-sectional views and/or plan views, which are ideal exemplary views of the present invention. In the drawings, the thicknesses of films and regions are exaggerated for effective explanation of technical content. Accordingly, the shape of the illustrated drawings may be modified due to manufacturing techniques and/or tolerances. Therefore, embodiments of the present invention are not limited to the specific shape shown, but also include changes in the shape generated according to the manufacturing process. For example, a region shown at right angles may be rounded or have a predetermined curvature. Accordingly, the regions illustrated in the drawings have attributes, and the shapes of the regions illustrated in the drawings are intended to illustrate a specific shape of a region of a device and are not intended to limit the scope of the invention. Although terms such as first and second are used to describe various elements in various embodiments of the present specification, these elements should not be limited by these terms. These terms are only used to distinguish one component from another. Embodiments described and illustrated herein also include complementary embodiments thereof.

Terminology used herein is for describing the embodiments and is not intended to limit the present invention. In this specification, singular forms also include plural forms unless specifically stated otherwise in a phrase. The terms 'comprises' and/or 'comprising' used in the specification do not exclude the presence or addition of one or more other elements.

In describing the specific embodiments below, several specific contents are prepared to more specifically describe the invention and aid understanding. However, readers who have knowledge in this field to the extent that they can understand the present invention can recognize that it can be used without these various specific details. In some cases, it is mentioned in advance that parts that are commonly known in describing the invention and are not greatly related to the invention are not described in order to prevent confusion for no particular reason in explaining the present invention.

Hereinafter, a fine dust prediction, ventilation control system and method using artificial intelligence transfer learning according to an embodiment of the present invention will be described. 1 shows a block diagram of a fine dust prediction system using artificial intelligence transfer learning according to an embodiment of the present invention.

First, artificial intelligence transfer learning applied in an embodiment of the present invention will be described. Figure 2 shows a schematic diagram of the basic principle of transfer learning. Transfer learning is a framework that can solve new but related tasks by leveraging previously learned knowledge. As shown in Fig. 2, transfer learning facilitates predictive modeling composed of various data patterns in a current domain by acquiring knowledge from data in a relevant domain. This mitigates the importance of frameworks in many real-world situations where obtaining significant amounts of data is time consuming, expensive or unavailable.

Frameworks are formally divided into two basic concepts: sources and targets. Domain D consists of two components given by D = {X, P(X)}. where X = {x1,., xn} is the n-dimensional feature space, and (P(X)) is the marginal probability distribution of X. Similarly, given a particular domain, T is defined as T={Y, P(Y=X)}. where Y is the target label and P(Y=X) is the conditional probability of Y given X. In the above relation, given a source domain D _s and learning task T _S , transfer learning with a target domain D _T and learning task _{T T} is learning of D _s and T _s if D _S ‡ D _T or T _S ‡ T _T It aims to learn the target conditional probability distribution using prior knowledge.

Among the methods of utilizing transfer learning, the recent success of machine learning has suggested a new category called network-based transfer learning. This method solves the biggest problem (i.e., insufficient training data) when developing complex deep learning models by reusing partial networks trained in the source domain to facilitate learning in the target domain.

The first strategy uses a pretrained model only for weight initialization. In these scenarios, both the feature extractor and classifier layers are trained, and final adjustments are made by fine-tuning the network in the target domain. In the second case, the feature extraction layer of the pre-trained model is completely fixed and the classifier layer is trained. This helps the model in domain adaptation while remembering previously learned tasks. Also, the final weights of the network are fine-tuned in the target domain. In the second approach, much less computational time is required because the network size is greatly reduced.

The motivation for using deep neural networks is to learn higher levels of abstraction by increasing the network depth. Contrary to intuition, the learning error starts to increase instead of decrease as the deep neural network converges, and this problem is called the network degradation problem. Deep ResNet is formulated as a series of basic blocks called residual blocks, which aim to learn the desired mapping H(x) by utilizing the residual function F(x). So instead of learning the mapping of H(x), the residual block has a different learning target and the network learns an alternative residual mapping given by the residual function F(x). This leads to the reformulation of the objective function shown in equations 1 and 2 below.

[Equation 1]

[Equation 2]

Figure 3 shows the building block matching of Deep ResNet (deep residual network). As shown in Fig. 3, skip concatenation is implemented in the basic block to skip other transforms F(x) to facilitate residual learning. For stochastic gradient-descent update calculations, using skip connections in a computation graph where gradient flows are defined allows gradients to be directly back-propagated to the bottom layer. Therefore, the skip connection guarantees that the gradient flow path of ResNet is much shorter than the depth of the network. This allows stacking the basic blocks in the network to increase modeling capacity without the vanishing gradient problem that can be lost during normal convolution in deeper networks.

The structure of the residual block consists of a series of convolutional layers, batch normalization layers, and activation functions. A convolution layer is defined by a set of filters and a convolution kernel core that performs convolution operations on the entire input. The feature map of the previous layer is convolved by the kernel function and the output is generated by the nonlinear activation function. Considering the functional map of the residual block in layer t denoted by u _t , the transformation from u _t-1 to u _t is given by Equation 3 below.

[Equation 3]

where f is the skip connection weight and w is the convolution kernel.

Hereinafter, embodiments of the present invention will be described. Given the MOE's emphasis on maintaining healthy IAQ levels, a remote monitoring system (TMS) was installed in SMS to track air quality levels. The TMS measures several air pollutants and metrological variables at 10-minute intervals, including PM ₁₀ , PM _2.5 , CO ₂ , temperature and humidity. The concentration of fine dust at the subway platform and outdoors was measured with a device based on the continuous beta line adsorption method, and the specifications of the measuring device are shown in Table 1. In addition, train service starts from the subway station from 5:00 am to 12:00 am, and runs every 3 minutes during rush hour and every 5 to 6 minutes at other times.

[Table 1] Air quality meter specifications

In the embodiment of the present invention, the platforms and outdoor locations of four subway stations (C, E3, E4, and D) of SMS Line 2 were measured from May 1 to 13, 2018. Figure 4a shows the change of each measured variable expressed as a time series at the E4-subway station. Figure 4b shows the location of the target subway station on line 2 of the SMS. In addition, subway specifications and basic measurement statistics are shown in <Table 2>.

[Table 2] - PM2.5 and PM10 specifications and basic statistics of target subway stations

FIG. 5A is C-subway station, FIG. 5B is E3-subway station, FIG. 5C is E4-subway station, and FIG. 5D is outdoor/indoor PM ₁₀ and PM _2.5 measurement distribution graphs of D-subway stations.

That is, FIG. 5 shows changes in PM ₁₀ and PM _2.5 concentrations of outdoor (left) and platform (right) during the measurement period at each observation station. Due to the proximity of these stations and the same measurement period, similar outdoor PM distribution trends were observed for all stations. On the other hand, the daytime behavior of PM can be seen at the platform, and the train frequency increases at the platform with the highest concentration in the morning and evening rush hours when passengers are crowded. Platform PM concentration is influenced by outdoor factors and becomes higher when outdoor conditions are poor. Despite the similar variations of PM, the unique characteristics of each subway station, such as shape, structure and passenger flow, result in slightly different IAQ dynamics at each station. Preliminary analyzes should be performed to determine statistical correlations between the IAQ data, considering the influence of multiple variables on platform PM _2.5 . In addition, the spatial correlation is analyzed by analyzing the PM distribution of neighboring stations.

First, the feature selection process is performed by calculating the Pearson correlation coefficient (PCC) between the available variables and the platform PM _2.5 concentration to determine the influencing variables and excluding redundant variables. Figure 6 shows the statistical correlation analysis of measurement variables for function selection and platform PM2.5. Changes in PM2.5 at subway platforms are influenced by several indoor and outdoor variables. Thus, it forms a complex dynamic that is not defined by a single variable. The PCC values show that the platform PM _2.5 has a relatively high positive correlation with the OPM ₁₀ and OPM _2.5 concentrations, due to outdoor contaminant penetration into the platform. The very high correlation between PM _2.5 and PM ₁₀ concentrations suggests that they originate from the same source. Platform CO ₂ levels are influenced by high train frequency and passenger density during morning and evening rush hours.

Train frequency, in turn, leads to PM2.5 particles caused by rail-wheel-brake wear and resuspension of particles on subway platforms. This platform CO ₂ level information can reflect rush hour and provide unique pattern information for PM _2.5 time series data. Temperature and humidity did not appear to have a significant effect on the platform PM _2.5 concentration, which is similar to previous studies. We also selected influence variables with PCC scores greater than 0.5 (i.e., moderate to high correlation) to reduce redundant input to the prediction framework. Therefore, based on outdoor PCC analysis, platform PM ₁₀ and platform PM _2.5 with CO ₂ were selected to predict future platform PM _2.5 concentrations.

In addition, a spatial correlation analysis was performed between the outdoor and indoor fine dust concentrations of the four observation stations. In particular, the correlation between OPM concentrations at nearby stations was 0.61 to 0.95. This suggests that the OPMs of these subway stations originate from similar sources. In addition, the OPM ₁₀ and OPM _2.5 dependencies of stations on platforms PM ₁₀ and PM _2.5 of other stations ranged from 0.38 to 0.68 and 0.38 to 0.69, respectively. OPM's PCC with neighboring stations decreased with increasing distance, except for the correlation between points C and E4, which was still high despite increasing distance. These strong spatial correlations between PM concentrations suggest that a single model can provide acceptable performance in predicting PM _2.5 concentrations. However, the individual dynamics of each location can affect the generalization performance of the model.

Hereinafter, the prediction framework for the platform CIAI (PM2.5) will be described. In a subway station, the ventilation system draws in fresh air from the outside to regulate the IAQ, and there is time required for the ventilation system to disperse the current PM level. Also in ventilation control, an operating period of 1 hour is typically implemented. A sequence prediction approach that tracks early detection of several PM _2.5 concentrations at hourly resolution is more viable than point-to-point prediction, because in our example, measurements are made in 10-minute increments. Therefore, the moving window concept was used to increase the validity and applicability of the time series prediction model. For example, the next 06 (1 hour) Platform PM _2.5 value can be predicted by looking at the past 2 hour observations of the variable selected to identify the temporal pattern of the neighborhood data. The size of the moving window is defined by past measurements to look back on and future observations to predict. In addition, in the embodiment of the present invention, a neural network model was developed using the Keras deep learning package, and the detailed algorithm is as follows.

Figure 7 shows the framework of ResNet: (a) convolution and ID block with activation function, (b) the overall structure of ResNet. The ResNet structure proposed to effectively learn features from IAQ time series data is a modified version of the ResNet-50 structure presented in FIG. The input layer is in the form of a two-dimensional matrix, and its size is defined by 5 input variables and 12 past observations given by a moving window. Here, each basic convolution block consists of a stack of three convolutional layers, batch normalization, an activation function, and a shortcut connection. A dropout layer was also added to the convolution block as a regularization method to avoid overfitting.

7(a) shows convolutional and identity blocks of a ResNet structure. Three pairs of convolutional, identity blocks are stacked in the main network structure, and the overall architecture of the ResNet model is shown in Fig. 7(b). Considering the zero gradient due to the “dying ReLU” problem, various activation functions (i.e. ReLU, SeLU and ELU) were used in different layers of the convolution block to improve the training process of the network. The identity and convolution blocks are followed by a fully connected layer with 64 neurons and an output layer with 6 neurons representing the future sequence of PM _2.5 CIAI on the metro platform. In this way, a 24-layer deep ResNet structure was developed. The hyperparameters are:

-Loss function: mean squared error (mse)

- Optimizer: Adam

- Batch Size: 18 (3 h)

- Epochs: 100

- Learning rate: dynamic (rate reducing factor 0.01 and patience of 10 epochs)

- Dropout rate: 0.2 (convolution block)

Hereinafter, a memory gate recurrent neural network according to an embodiment of the present invention will be described. In the embodiment of the present invention, modifications of repeating units represented by RNNs (eg, LSTM and GRU) are applied. For comparison, we implemented a stacked RNN (S-RNN) structure, specifically a stacked LSTM (S-LSTM) and a stacked GRU (S-GRU) structure. In S-RNN, several hidden recurrent layers overlap each other. The output of one recurrent layer becomes the input of the next layer. Here, we constructed two types of S-RNN structures, one by stacking two recurrent layers (S-RNN-2), and the other by stacking 10 recurrent layers (S-RNN-10) for deeper to develop a network. The input layers of all RNNs were identical to the ResNet structure, and the rest of the architecture was defined as 1) 2/10 hidden layers with 64 neurons to store the hidden representations and 2) output layers with 6 neurons representing the prediction order. . The rest of the hyperparameters are given as:

-Loss function: mse

- Optimizer: Adam

- Batch Size: 18 (3 h)

- Epochs: 100

- Learning rate: dynamic (rate reducing factor 0.01 and patience of 10 epochs)

- Dropout rate: 0.2

Describing the setting and implementation for transfer learning, the embodiment of the present invention focuses on utilizing transfer learning to solve the subway IAQ data shortage problem. The transfer learning framework works under the assumption that there is sufficient monitoring data (ie source domain) to develop a reliable and acceptable predictive model for a subway station to transfer to another subway environment (ie target domain). The main goal of the present invention is to improve prediction and ventilation control performance when training data is lacking due to lack of advanced monitoring systems or newly developed subway stations.

8 is a graphical representation of a training process for a transfer learning framework in accordance with an embodiment of the present invention. The transfer learning infrastructure was developed by adopting the network-based training methodology as shown in FIG. To account for the fine-tuning of the transition model, we set the network parameters of the feature extractor to perform operations on the source dataset (X _s , Y _s ) and the classifier on the target dataset (X _T , Y _T ) to θ _S1 and θ respectively. _S2 and θ _T1 and θ _T2 as parameters for the task.

First, a standalone model was pretrained at a subway station with sufficient data, and then the feature extractor parameters of the pretrained model were frozen. A frozen layer indicates that the layer's weights remained unchanged when the model was assigned to perform another task. The rest of the classifier layer was fine-tuned for data from the new subway station. From this point of view, the model memorizes the PM distribution from enough data by freezing the initial layer and applies it to the dynamics of the next subway station with minimal loss during refinement. In addition, the optimal parameter θ obtained through fine tuning is expressed by Equation 4.

[Equation 4]

The Comprehensive Indoor Air Quality Index (CIAI) is designed as a tool for the general public to understand IAQ status and associated health risks. The CIAI categorizes the health risks of several indoor air pollutants into six levels of concern, with values of unhealthy, unhealthy, very unhealthy, and dangerous conditions for good, moderate, and sensitive groups.

The CIAI for each indoor air pollutant can be calculated by the given Equation 5.

[Equation 5]

where BP and I are the concentration and exponential breakpoints for each level of health concern. LO and HI represent the upper and lower bounds of each index. C _p also identifies the concentration of a particular contaminant.

The effectiveness of the forecasting methodology was evaluated using four performance indicators. These include root mean square error (RMSE), mean absolute error (MAE), mean absolute percentage error (MAPE) and coefficient of determination (R ² ). These measurement items are provided as in Equations 6 to 9 below.

[Equation 6]

[Equation 7]

[Equation 8]

[Equation 9]

where n is the number of test observations, y _i is the observed CIAI of PM _2.5 , and yi* is the predicted CIAI value dl. Lower error metric values and higher R ² values indicate higher accuracy and better prediction performance.

Hereinafter, a prediction methodology proposed in an embodiment of the present invention will be described. 9 shows a schematic diagram of a methodology performed for sequence prediction of PM _2.5 CIAI at a subway platform according to an embodiment of the present invention. The prediction framework according to an embodiment of the present invention is divided into three main steps.

First, in order to directly quantify the future information of PM _2.5 in terms of related health risks, the input variables selected in the preliminary data analysis mentioned above were converted into CIAI. Quantification of CIAI also provides an understanding of the contribution of variables in IAQ dynamics. The moving window concept was then used to predict a series of PM _2.5 concentrations tracked by CIAI, ensuring a feasible forecasting approach. A moving window was constructed by selecting 12 past observations (2 h) to identify complex patterns in the neighborhood data and predicted 6 data points (1 h).

In the second step, the data set is divided into a training (80%) set and a test (20%) set. Multiple sequence prediction was performed with RNN and CNN algorithms at four different subway stations to select the most suitable structure. RNN and CNN architectures include implementations of S-LSTM-2, S0LSTM-10, S-GRU-2, S-GRU-10 and ResNet-24, respectively. A multiple linear regression (MLR) model is also included for comparison with machine learning models. The structural performance of each station was compared using the aforementioned metrics.

Third, when an appropriate framework is selected, transfer learning substructures are combined with predictive algorithms to solve the problem of insufficient data. To thoroughly examine the performance of the proposed transfer learning-based methodology, we simulate four cases by assuming that 20%, 40%, 60% and 80% of the data are available for training, while keeping the test set the same. The transfer learning model was developed through pre-training on data from the base station (source domain) and fine-tuned on the data available at the new station (target domain).

Finally, the generalization ability of the model was considered. The intuition behind this procedure is to learn the PM2.5 distribution in various IAQ work environments without losing past learning experience. Thus, it learns the complexity of IAQ dynamics in a wider range. We checked the generalization performance of the transmitted network for stand-alone models.

Subway ventilation control systems play an important role in reducing the health risk exposure of commuters on subway platforms and the energy demand of subway stations. The reliability of future air quality data has a significant impact on the performance of ventilation systems. Misinformation can lead to excessive energy demand by higher inverter frequencies (60 Hz) or hazardous IAQ environments at lower inverter frequencies (20 Hz) to commuters. Therefore, the purpose of the ventilation control system evaluation using future air quality information is to evaluate the control meaning of a stand-alone model with insufficient IAQ data on ventilation performance compared to a transfer learning-based framework.

Ventilation control responses according to the predicted data of the stand-alone model and the previous model were compared. Two scenarios are considered here. (1) 40% IAQ data available on the new station; (2) 80% available data. In this regard, the IAQ system for Line 2 of the SMS should be accepted to evaluate the behavior of the ventilation system in this scenario. Loy-Benitez et al. proposed a mathematical model for the metro IAQ process for SMS's Line 2 that transforms the relationship between ventilation demand and _PM10 concentration given by the transfer function (Gp) in Equation 10.

[Equation 10]

If the PM _2.5 /PM ₁₀ mass ratio of Seoul Subway Line 2 is calculated as 0.80, the transfer function relating PM _2.5 concentration and ventilation energy demand can be expressed by the prediction error minimization (PEM) method as shown in Equation 11. there is.

[Equation 11]

10 is a schematic representation of a transfer learning-based framework proposed for sequential prediction and ventilation control of the platform PM _2.5 CIAI level according to an embodiment of the present invention. The response of the ventilation system was evaluated according to the transfer function (GPM _2.5 ) based on the health risk of commuters assigned by CIAI and the ventilation energy demand of the control system. In addition, the ventilation energy consumption is estimated using the 3rd order polynomial of Equation 12 below.

[Equation 12]

Let us explain the performance prediction for direct training. The PM _2.5 sequence prediction provided by the subway platform moving window was performed for early detection at the CIAI level by considering the dynamics of four subway stations in SMS. In the first phase of the experiment, the performance of six standalone network architectures was compared. These networks include implementations of ResNet structures based on MLR, LSTM, GRU and CNN. LSTM and GRU networks include shallow and deep variants such as S-LSTM-2, S-LSTM-10, S-GRU-2 and S-GRU-10. In addition, a predictive model was trained for each station according to the aforementioned detailed parameters. For direct training (no transfer learning), 80% of the data was used for model training and 20% was used as a test set for evaluation.

Table 3 - Platform PM2.5 CIAI sequences predicting the performance of RNN and ResNet structures in direct training

Table 3 shows the metric scores of the neural network for each station. Based on the performance results, meaningful observations can be made. The results show that the performance of the MLR was relatively low compared to the RNN network, and the existing regression model could not identify the dynamic temporal pattern of PM _2.5 concentration. The RMSE values generated by the same transformation of the GRU and LSTM structures were very similar. However, as the depth of the network increased, the performance of the S-LSTM-10 and S-GRU-10 networks decreased compared to the shallow variant. The RMSE values of S-LSTM-10 and S-GRU-10 decreased by an average of 16.19% and 8.34%, respectively. In addition, R ² indicates that the capture of the described variances by S-LSTM-10 was greatly reduced by an average of 50.02% compared to the shallow variances. This demonstrates the vanishing gradient problem when training deeper neural networks.

On the other hand, skip connection in 24-layer deep ResNet enables neural network learning without the problem of vanishing gradients and showed the lowest evaluation score. Based on RMSE metrics, the proposed ResNet structure outperforms S-LSTM-2 and S-GRU-2 by averages of 10.61% and 8.91%, respectively. Also, R ² improved by 27.90% and 42.84%. Clearly, the predictive performance of the model varies from station to station, with the E-3 subway showing the lowest RMSE (11.8826) and the E-4 subway showing the highest R ² value (0.78). The difference in the evaluation results was influenced by the unique PM distribution, which varied according to several factors such as passenger flow, shape (island or plane) and depth of the subway platform.

11 is a sequence prediction of platform PM _2.5 CIAI by moving window in E3-subway according to an embodiment of the present invention: (a) MLR, (b) S-LSTM2, (c) S-GRU2, (d) S- For LSTM10, (e) S-GRU10 and (f) ResNet, the shaded area represents the 95% confidence interval. It can be observed here that all network architectures aim to capture the temporal variation of the _PM2 distribution given by a moving window. Also, the 95% confidence interval shows that the nonlinear sequence of PM _2.5 is predicted with acceptable performance. However, LSTM and GRU networks showed slightly higher standard deviations compared to the ResNet structure. This is interpreted as a fluctuating response to small changes in PM _2.5 concentration. This result is also indicated by a low R ₂ highlighting the false capture of the explained variance.

On the other hand, the values predicted by ResNet did not show large fluctuations, and the 95% confidence limits were very close to the observed concentrations. Moreover, the variance by the RNN model generated false positives, overestimating or underestimating the predicted values when compared to the observed CIAI values. This resulted in higher prediction errors moving from one CIAI domain to another. These fluctuations may also result in higher ventilation energy demands or increased health risks for commuters due to lower air supply.

It should also be noted that several other compounds such as VOCs, carbonaceous compounds, crustal materials and heavy metals can also affect platform PM _2.5 . However, some of the contributing complexes are self-captured by PM _2.5 measurements. Therefore, considering the influencing IAQ variables and historical PM _2.5 measurements as inputs, the neural network model aims to learn the unique physicochemical principles controlling platform PM _2.5 concentrations in a time-series format. Nevertheless, initial experiments using metric scores and temporal trends in the PM _2.5 CIAI domain showed that the ResNet structure performed relatively well compared to MLR, LSTM and GRU networks. Therefore, it was chosen to develop a transfer learning framework for further experiments.

Hereinafter, prediction performance results for transfer learning will be described. A transfer learning framework was developed after selecting an appropriate predictive model. To demonstrate the effectiveness of transfer learning, we considered the limited data available for training and the experimental phase is divided into two parts. First, the ResNet structure was pre-trained under direct training, and second, the baseline model was fine-tuned based on the insufficient data of the new station. E3 subway station was selected to develop the baseline model because it was located between other subway stations and showed a relatively high PCC score compared to other subway stations. In addition, the spatially connected metro stations emphasized similarities between the source and target domains. The training samples prepared for the transfer learning experiments are shown in Table 4.

[Table 4] - Data sample prepared for transfer learning experiment

The number of fixed layers of the baseline model was determined before fine-tuning the model at other stations. The intuition for this is that a large number of layers can be used for training and when a new station is overfitted, the model may not remember enough knowledge of the source data. On the other hand, if the number of frozen layers is too large, the model may not be able to update enough parameters to adapt to the target station and performance may suffer. To account for the impact of the limited IAQ data, we first set the data availability for training of the new station to 40%, while keeping the test set the same for comparison.

To demonstrate the effect of frozen layers, the layers of the baseline ResNet pretrained at E-3 station were frozen from 20% to 70% and then the model was transferred to station E4. Table 5 reflects the prediction results of transfer learning-based ResNet (TL-ResNet) evaluated with performance metrics for different numbers of freezing layers. As a result of the analysis, the prediction performance of TL-ResNet starts to improve as the number of fixed layers is increased first, and shows the highest performance at 40% fixed layers. Increasing the number of fixed layers tilts the network more towards the base station and reduces performance. Based on this analysis, the optimal number of freezing layers was found to be 40%, and this number was chosen to further develop transfer learning-based models for different stations. After finding the optimal number of frozen layers, we measured the predictive capabilities of TL-ResNet over stand-alone ResNets. The baseline ResNet was first trained with data from E3-subway station and retrained from other subway stations to account for the lack of data. Table 6 shows the performance of standalone ResNet and TL-ResNet at each station evaluated by the test set. Standalone ResNets appear to have large prediction errors on limited data. In contrast, transfer learning frameworks amplified sequence prediction and provided significant improvements over standalone architectures. Considering the availability of 40% training data, the RMSE performance of E4-Metro, C-Metro, and D-Metro increased significantly by 33.97%, 28.84%, and 22.80%, respectively, for an average increase of 28.54%. Also, R ² of C-subway improved significantly from 0.1222 to 0.5085, with the other two stations showing an average increase of 58.14%.

[Table 5] - Effect of fixed layer on sequence prediction performance of TL-ResNet

[Table 6] - Multi-sequence prediction using standalone ResNet and TL-ResNet (with 40% usable data for transmitted subway stations)

CIAI temporal trends identified at each subway station are compared and shown in FIG. 12 . The 95% confidence interval indicates that the standalone ResNet cannot capture the unique changes of the subway IAQ system in limited data scenarios, especially for moderately aggravated IAQ conditions. In contrast, TL-ResNet captures the dynamics of subway IAQ even when available monitoring data is scarce. It also shows a significant improvement by predicting the platform PM _2.5 CIAI domain with acceptable performance.

Figure 13 shows RMSE scores and performance improvement graphs of standalone ResNet and TL-ResNet on E4, C and D-subway using different data availability scenarios. Figure 13 shows the RMSE scores by standalone ResNet and TL-ResNet with different cutoff points in the available training data. In particular, the RMSE value decreased after applying the transfer learning-based algorithm to all subway stations. The performance improvement was greatest at 20% of the available data, with an average improvement of 37.58%. As a result, the value of transfer learning gradually decreased as the available training data increased, and an average performance increase of 11.30% was observed at 80% data availability. This result is expected because as more training data is included, the standalone model begins to identify IAQ dynamics.

Finally, transfer learning was applied in a series of steps to improve the generalization and adaptability of the TL-ResNet structure. The baseline model was first trained on the E-3 subway and then transferred over the IAQ data of the E-4 and C-metro stations. To develop a generalized model, we considered complete training data available from all stations, as we aim to learn from different IAQ environments. Compared to base stations, the prediction performance of standalone ResNets decreased at nearby stations because they constitute slightly different IAQ dynamics. On the other hand, after two-step learning, the RMSE performance of TL-ResNet increased by 11.08% and 14.58% in E4 and C-subway, respectively. However, the performance at the base station decreased slightly by 4.86% as it moved from a specific area to a more general area. Also, at the 4th station (D-subway where learning was not performed), R ² improved from 0.66 to 0.75.

The results show the effectiveness of transfer learning in helping TL-ResNet effectively identify temporal changes in PM _2.5 while learning in a mixture of IAQ environments remembers previously learned dynamic IAQ information. Thus, TL-ResNet represents a reliable subway air quality prediction framework that integrates conflicting air quality relationships and scarce monitoring data. In addition, the generalized network architecture is not limited to specific IAQ conditions, but can identify various IAQ environments to quantify the health risk areas of subway commuters.

The aforementioned experiments focused on predicting the performance of standalone and TL-ResNet frameworks. Hereinafter, the meaning of the predicted IAQ information for the ventilation system to control the future PM _2.5 concentration will be explained. In addition, the consequences of misinterpreting the IAQ environment for the ventilation control system were evaluated.

To this end, experiments were conducted at the E4 subway station during the last two days of the test set (288 samples). The sequence prediction results by standalone and TL-ResNet were utilized in 40% and 80% data availability scenarios. The control systems for each scenario were compared in terms of the platform's CIAI level and energy demand. 14(a) shows the experimental data on the E4 subway platform for scenario 1. The black line represents the actual measured value of PM _2.5 on the subway platform and the dotted line represents the sequence prediction by standalone ResNet and TL-ResNet.

Figure 14(b) shows the response of the ventilation system in the subway platform and CIAI level by standalone ResNet for 40% available IAQ data. The red dotted line represents the upper limit of the inverter frequency and the yellow dotted line represents the breakpoint at a good to moderate IAQ level.

Ventilation control performance based on the two indicators is shown in Table 7. Here, the underestimation of the IAQ environment by the standalone ResNet allowed the ventilation system to interpret this erroneous information as a good IAQ level. In addition, the inverter frequency did not increase to cope with the deteriorating IAQ condition as high air volume was required due to persistent misinformation about the IAQ environment. The ventilator system failed to identify the correct IAQ condition, resulting in lower IAQ levels, with 282 points exceeding the median CIAI level breakpoint. Platform PM _2.5 concentrations in this time interval averaged 20.81 μg/m ³ with a moderate level of concern, and the lower air supply resulted in energy consumption of 1428.62 kWh/day.

Table 7 - Ventilation control system performance for 40% and 80% available IAQ data

In contrast, the predicted data generated by TL-ResNet closely matched the values measured at actual subway platforms. Thus, the ventilation system can identify IAQ conditions that are deteriorating. To cope with this, the inverter frequency was increased as shown in FIG. 14(c). As a result, more air supply was achieved at the subway platform, and the platform PM _2.5 was maintained at a good IAQ level with an average concentration of 14.73 μg/m ³ . In this case, 129 of the 288 observations crossed the median CIAI breakpoint. As a result, the average value of PM _2.5 in the platform improved by 29.21% compared to the single system. It can be seen here that the ventilation system attempted to balance energy consumption with high air delivery in deteriorating IAQ conditions and low inverter frequency in healthy conditions. During this period, high inverter frequency increased energy demand to 1917.83 kWh/day. Additionally, high air transfer on subway platforms has the potential to expose commuters and subway workers to moderate PM _2.5 levels for 26.5 hours.

It should be noted that the energy consumption by TL-ResNet is higher than that of a stand-alone system in both scenarios. The reason is that the experimental period applied to the embodiment of the present invention showed a high PM2.5 concentration during the last two days. Therefore, a high energy demand response would have been produced by the ventilation system when the control system followed observations close to the actual IAQ data, and vice versa would have produced a lower energy demand. It is known that there is a conflicting relationship between air quality level and ventilation energy demand in indoor building spaces, which are generally affected by many uncontrollable factors such as outdoor dust storms, high commuter flow, and sensor failures. A balance between them is therefore often required. Nonetheless, monitoring for IAQ may be insufficient given the sensor's reliability, operation and installation in a hostile building environment. Therefore, it can be seen that transfer learning-based methods can provide a sustainable solution and provide significant improvements over standard methods to solve the problem of insufficient data in IAQ monitoring and control.

As mentioned above, in the embodiment of the present invention, we proposed a transfer learning-based deep ResNet framework for sequential prediction of PM _2.5 -related health risk levels in subway platforms. In an embodiment of the present invention, the ResNet architecture and transfer learning framework are integrated to provide reliable predictions in situations where data is insufficient. Experiments were conducted to demonstrate the influence of reliable and incorrect IAQ information on ventilation control systems in terms of PM _2.5 CIAI level and energy consumption. The remaining results based on the analysis and experiments according to the embodiments of the present invention are as follows.

- Compared to shallow RNN networks, deeper variants showed poor sequence prediction performance, highlighting the gradient vanishing problem in deep neural networks.

- Transfer learning frameworks can provide a sustainable solution to address data availability issues when developing acceptable predictive models for air quality forecasting.

- TL-ResNet can effectively improve IAQ predictions and regulate ventilation systems in scenarios with insufficient data to achieve sustainable CIAI levels in subway platforms.

In addition, the device and method described above are not limited to the configuration and method of the above-described embodiments, but all or part of each embodiment is selectively combined so that various modifications can be made. may be configured.

Claims (11)

In the fine dust prediction system,
a predictive model generator for constructing a predictive model based on transfer learning; and
A fine dust prediction system using artificial intelligence transfer learning, comprising: a fine dust prediction unit for predicting the concentration of fine dust after a specific time based on the prediction model built through the prediction model generation unit.
According to claim 1,
The prediction model is a fine dust prediction system using artificial intelligence transfer learning, characterized in that the transfer learning-based ResNet prediction model.
According to claim 2,
The specific time is 40 minutes to 100 minutes, and the fine dust data is PM _2.5 data, characterized in that the fine dust prediction system using artificial intelligence transfer learning.
According to claim 3,
The ResNet prediction model is a fine dust prediction system using artificial intelligence transfer learning, characterized in that for applying source data in a space other than the space to be predicted.
According to claim 4,
Using the ResNet prediction model, the prediction system is trained to predict the PM _2.5 concentration after a specific time in the prediction target space even when the learning data of the target space to be predicted is insufficient by using the source data of other spaces. A fine dust prediction system using artificial intelligence transfer learning.
According to claim 5,
The ResNet structure is a fine dust prediction system using artificial intelligence transfer learning, characterized in that it includes a convolution block, an identity block, a fully connected layer, and an output layer.
According to claim 6,
The ResNet prediction model is derived through pre-training on source data, and fine dust prediction system using artificial intelligence transfer learning, characterized in that it is fine-tuned for data available in the target space.
According to claim 7,
The ResNet prediction model is a fine dust prediction system using artificial intelligence transfer learning, characterized in that the initial layer of the source data is fixed, the PM distribution is stored in sufficient data, and the loss during fine-tuning is minimized and applied to the target space.
According to claim 8,
The ResNet prediction model fixes the feature extractor parameters of the pre-trained model after pre-training the source data, and the fixed layer keeps the weight of the layer unchanged when the predictive model is assigned to perform another task. A fine dust prediction system using artificial intelligence transfer learning, characterized in that the remaining classifier layers are fine-tuned for the target data of the target space.
According to claim 9,
The fixed layer is a fine dust prediction system using artificial intelligence transfer learning, characterized in that 35 to 45%.
A program that is readable by a computer and causes the fine dust prediction system to be executed using artificial intelligence transfer learning according to any one of claims 1 to 10.