|تعداد مشاهده مقاله||106,241,866|
|تعداد دریافت فایل اصل مقاله||83,141,705|
توسعه مدل ترکیبی برپایه شبکه عصبی جهت مدلسازی کیفیت هوای داخلی بازار تبریز به لحاظ ذرات معلق
|مقاله 8، دوره 45، شماره 3، آبان 1398، صفحه 499-512 اصل مقاله (1.76 M)|
|نوع مقاله: مقاله پژوهشی|
|شناسه دیجیتال (DOI): 10.22059/jes.2019.288082.1007917|
|ندا کفاش چرندابی* ؛ امیر غلامی|
|گروه نقشه برداری، دانشکده فنی مهندسی مرند، دانشگاه تبریز، تبریز، ایران|
|با توجه به سپری شدن ساعات متعددی از زندگی هر انسان در فضاهای بسته، بررسی کیفیت هوای داخلی اماکن بسیار ضروری است. آلاینده های مختلفی در فضاهای بسته وجود دارند که از بین آنها، ذرات معلق توجه زیادی را به علت تاثیرات فاجعه بار بر سیستم تنفسی و حتی سیستم گردش خون به خود جلب کرده است. در این تحقیق با جمع آوری داده های مورد نیاز از فضاهای داخلی بازار تبریز کیفیت هوای آن به لحاظ ذرات معلق برپایه مدلهای مختلف شبکه عصبی مورد بررسی قرار گرفت. طول معابر، عرض معابر، تراکم نسبی جمعیت، جنس کف، وجود یا عدم وجود تهویه، تنوع کاربری، فاصله از خیابان های مجاور، بلندی سقف، دما و رطوبت نسبی در 1081 موقعیت مختلف اخذ شد که 86 نقطه از آنها دارای داده مربوط به آلاینده های ذرات معلق بودند و به عنوان داده ورودی برای ساخت شبکه بکار رفتند. پس از مقایسه شبکه های مختلف با پارامترهای گوناگون شبکه بهینه ساخته شد. سپس جهت افزایش کارایی مدل پیشنهادی، شبکه ساخته شده با طبقه بند ترکیبی ادغام گردید. نتایج به دست آمده از مدل پیشنهادی تحقیق به صحت 97.67 درصد برای پیش بینی کلاس آلاینده ها رسیده است.|
|آلودگی هوا؛ ذرات معلق؛ بازار تبریز؛ شبکه عصبی؛ طبقه بند ترکیبی|
|عنوان مقاله [English]|
|A Hybrid Model Based on Neural Network for Indoor Air Quality Modeling of Tabriz Bazaar in Terms of Particulate Matter|
|Neda Kaffash Charandabi؛ Amir Gholami|
|Department of Geomatics,, Marand Technical Colleague, Tabriz University, Tabriz, Iran|
|The term of neural network has evolved to encompass a large class of models and learning methods. They interpret input raw data through a kind of machine perception, labeling or clustering models. They help to group unlabeled data according to similarities among the example inputs, and they classify data when they have a labeled dataset to train on. |
The neural networks consist of many layers. The layers are made of nodes. A node is just a place where computation happens, loosely patterned on a neuron in the human brain, which fires when it encounters sufficient stimuli. A node combines input from the data with a set of coefficients, or weights, which either amplify or dampen that input, thereby assigning significance to inputs with regard to the task the algorithm is trying to learn. These input-weight products are summed and then the sum is passed through a node’s so-called activation function, to determine whether and to what extent that signal should progress further through the network to affect the ultimate outcome, say, an act of classification. If the signals pass through, the neuron has been “activated.”
A node layer is a row of those neuron-like switches that turn on or off as the input is fed through the net. Each layer’s output is simultaneously the subsequent layer’s input, starting from an initial input layer receiving your data. Pairing the model’s adjustable weights with input features is how we assign significance to those features with regard to how the neural network classifies and clusters input.
Ensemble methods aim at improving the predictive performance of a given statistical learning or model fitting technique. The general principle of ensemble methods is to construct a linear combination of some model fitting method, instead of using a single fit of the method. Ensemble classifier has two common methods: Bagging and boosting that they are non-Bayesian procedures. They can be used to improve the accuracy of Classification & Regression Trees.
Boosting and bagging are two ensemble models capable of squeezing additional predictive accuracy out of classification algorithms. When using either method, careful tuning of the hyper-parameters should be done to find the best balance of model flexibility, efficiency & predictive improvement.
Bagging uses a simple approach that shows up in statistical analyses again and again — improve the estimate of one by combining the estimates of many. Bagging constructs n classification trees using bootstrap sampling of the training data and then combines their predictions to produce a final meta-prediction.
In this way, boosting creates successive base classifiers that are told to place greater emphasis on the misclassified samples from the training data. Like bagging, the results from all boosting base classifiers are aggregated to produce a meta-prediction. Compared to bagging, the accuracy of the boosting ensemble improves rapidly with the number of base estimators. Boosting method has many different algorithms that Least Squared Boosting (LSBoost) model was used to improve prediction model in this paper.
Discussion of Results & Conclusions:
In the first step, for implementing and evaluating the research’s suggested method, Tabriz bazaar was chosen as the world’s biggest, most important and most complicated covered spaces. As the most important reasons for this choice the large dimensions of this historical bazaar, its complicated structure, presence of a wide range of tradespeople and different occupiers in bazaar, citizens’ and tourists’ referral all year long can be mentioned.
Then different information related to Tabriz bazaar’s various passages and buildings including Passage’s length, passage’s width, proportional population density, surface material, presence or absence of ventilation, user’s diversity, distance from nearby streets, roof’s height, temperature and proportional humidity was recorded via field observation. Afterwards training data were collected for neural network. This was done in 86 points with different conditions with the help of a Japanese sensor named Pocket PM2.5 sensor which can evaluate PM2.5 and PM10 pollutants in the range of 0-999 (μg/m^3 ).
Collected data for both of the pollutants were classified based on the present standards and place’s conditions in four healthy, medium, warning and dangerous classes and was shown by one to four tags. Given that determining parameters like the number of hidden layers, the number of neurons, the number of replications, the kind of transferring and tutorial functions in neural networks’ utilization is very important and effective, to determine them optimally a code in matlab coding language was used that not only it checks all the feasible functions and the appropriate number for neurons and layers, but also it determines their best amounts. By making multilayer neural network it is possible to predict dispersion in various conditions of pollutants’ amounts. In this paper due to users’ requirement for noticing the pollutant’s class and increasing accuracy the Round operator was used for classificating the results to four classes one to four.
Due to the low accuracy of the results, ensemble classifier was used to improve the network. LSBoost model of Ensemble algorithm with the help of 100 boosted decision trees has obtained very appropriate results for this paper’s case study with a learning rate of one (η=1). for performing the neural network in training and network creation phase, data were divided into three sections of train, validation and test data that the way of mean squared error’s convergency and result’s accuracy were evaluated for each of the PM2.5 and PM10 pollutants. Efficient network was choosen by convegency’s speed, performance and accuracy.
Due to the implementation the best network for PM2.5 pollutant was determined by two hidden layers (first layer with 9 neurons and transfer function of poslin and second one with 5 neurons and transfer function of tansig) with traincgb’s train data and general accuracy of 97.67% and MSE of 0.385. Also for PM10 pollutant with a hidden layer containing 4 neurons, transfer function of tansig and train function of trainlm with the general accuracy of 97.67% and MSE of 0.2779 was obtained. Traincgb function of Conjugate gradient backprogation with Powell-Beale restarts and trainlm of Levenberg Marquadrat algorithm is a combination of Gauss-Newton and gradient descent which uses the ability of both algorithms and has a high speed in training.
It should be noted that in addition to the aforementioned networks, the Ensemble classification algorithm was run on this paper’s data which was resulted to accuracy of 67.44% for PM10 and to accuracy of 68.60% for PM2.5 that indicates the optimality of the compositional approaches used in this study. In other words by using combinational algorithm Ensemble and backpropagation multilayer neural network, the accuracy of results have been improved about 30%.
Then based on the collected data in different conditions, the training was done for PM2.5 pollutant with the help of Newff-Ensemble neural network and for PM10 with the help of Feedforward-Ensemble network and prediction was done for other 995 locations of bazaar and PM zoning map was prepared. Based on these maps it is clear that most of the center zones of Tabriz bazaar has appropriate air quality. Some sections of bazaar including bazaarche shotorban, bazaarche yakhchal, chaharsog sadegieh, bazaar dalaleh zan bozorg, raste bazaar sadegieh, karvansarayeh shazdeh bozorg, dalan khan and bazaar jambor ha were in 4th class because of particulate matters and attendance of people who suffer respiratory disorders. Based on collected data for this study and by checking the results of this paper’s proposed model, four criteria Passage’s length, proportional population density, presence of ventilation, user type have been much more effective in increasing particulate matters’ class . So due to the obtained results from this study, areas which are located in dangerous class, effective and serious solutions should be done. For example related to the construction of these areas, modern or classic ventilation systems can be installed to improve the quality of bazaar’s indoor air.
|Air pollution, Particulate Matter, Tabriz Bazaar, Neural Network, Ensemble methods|
تعداد مشاهده مقاله: 622
تعداد دریافت فایل اصل مقاله: 161