ARCHITECTURE

• PROBLEM DESPRICTION(專案描述)
• HIGHLIGHT(重點部分)
  • PRE-PROCESSING(資料預訓練)
  • ADJUST LEARNING RATE(調整學習係數)
  • FEATURE SELECTION(特徵值更換)
  • FEATURE SCALING(特徵縮放)
  • PERFORMANCE IMPORVEMENT(提升預測結果)
• LISTING(條列程式檔案)
• TEST AND RUN(程式的測試與執行)
• DISCUSSION(結果討論)

• Data source
  “環境保護署 2019 年桃園市平鎮區空氣品質監測資料”

• Choose Model
  

  • Let "y'" be the prediction PM2.5,
    "x" be the features of 9 hours sensor types (ex. CO, NO, PM10, PM2.5, Rainfalls),
    "w" be the weights of features
  • Function 1: y’ = w * x
  • Function 2: y’ = w1 * x + w2 * x^2

• Loss function
  RMSE (root-mean-square error)

• Gradient descent
  Vanilla (basic), SGDM, MBSGD, AdaGrad, RMSProp, Adam

• Flow chart:
  

• Using Python rather than excel
• Invalid value, null value
  filled with the mean of sensor types
• Split training dataset
  20 days/month
• Split testing dataset
  Remaining days/month (8~11 days)
• Data extraction
  Every 9 hours with 15 sensor type/hour be the features and predict the PM2.5 of 10th hour

• Model: Power of one
• Gradient: Vanilla g radient descent
• Iteration: 1,000 times

• Training set number: 4,521
• Validation set number: 1,131
• Iteration: 10,000 times
• Gradient: AdaGrad

• Z-Score (Standard Score): (x – μ) / σ
• Max-Min: x - min(x) / max(x) - min(x)

• Training set number: 5,652
• Testing set number: 1,446
• Iteration: 150,000 times

• DataPreprocession.py: Data cleaning, Dataset spliting
• TestDataPreprocessing.py: Feature extraction, Shuffle
• LinearRegression.py: Feature extraction, Shuffle, Normalization,Split validation set, Training, Prediction
• ReDesignModel_Seq.py: Same as LinearRegression.py but use "square function model"
• getHT.py: Raspberry Pi reads temperature and humidity datas from DHT11
• CSVDownloader_GoogleSheet.py: Automatically download datas frominternet google sheet
• getUbidotsData.gs: Get datas from Ubidots to Google sheet, "gs"means "Google Apps Script"

1. What's happened to adjusting learning rate?
   
   • If learning rate is not minimal enough, loss would not converge (even become greater than greater)
     
   • If learning rate is too small, loss convergence would be inefficient
   • Only learning rate just fit, loss convergence would be efficient
2. Why removing some features would improve the performance?
   • Maybe the other features effect on PM2.5 be not that fast
3. Why Z-Score is better in performance of convergence?
   • Z-Score is appropriate in unknowing maximum and minimum
4. How to improve model?
   1. Shuffle training and testing data
   2. Using AdaGrad rather than Vanilla gradient descent
   3. Feature only select PM10 and PM2.5
5. Final improvement of model:
   converge loss value descents 0.0225