
# Author: Stephen Situ
# This is a project that takes US Air Quality data and creates a regression model based on a deep learning neural network model with 3 dense layers
# to predict AQI (Air Quality Index) that ranges from 0 to 300 typically. 
# using tensorflow/keras in python. The data was split into training and testing data sets, hot encoded for the categorical variables,
# and Min/Max scaled. Using a basic deep learning/neural network model, we achieve a very low MAE of 0.0067.
# Original Dataset can be found here: https://www.kaggle.com/datasets/calebreigada/us-air-quality-1980present

# Import Libraries
import numpy as np
import pandas as pd


# Read CSV
air_quality = pd.read_csv('US_AQI.csv')
air_quality.head()


# Data Summary
air_quality.dtypes

CBSA Code                      int64
Date                          object
AQI                            int64
Category                      object
Defining Parameter            object
Number of Sites Reporting      int64
city_ascii                    object
state_id                      object
state_name                    object
lat                          float64
lng                          float64
population                     int64
density                        int64
timezone                      object
dtype: object


# check duplicate
duplicate_rows = air_quality[air_quality.duplicated(keep='last')]
print(duplicate_rows)

Empty DataFrame
Columns: [CBSA Code, Date, AQI, Category, Defining Parameter, Number of Sites Reporting, city_ascii, state_id, state_name, lat, lng, population, density, timezone]
Index: []


# check null
nan_rows  = air_quality[air_quality.isna().any(axis=1)]
print(nan_rows)

Empty DataFrame
Columns: [CBSA Code, Date, AQI, Category, Defining Parameter, Number of Sites Reporting, city_ascii, state_id, state_name, lat, lng, population, density, timezone]
Index: []


# Split Date column from string into three columns, month, day, year
date_split = air_quality["Date"].str.split("/", n = -1, expand = True)
air_quality["Month"]= date_split[0]
air_quality["Day"]= date_split[1]
air_quality["Year"]= date_split[2]
air_quality.info

<bound method DataFrame.info of          CBSA Code       Date  AQI  Category Defining Parameter  \
0            10140   1/1/2022   21      Good              PM2.5   
1            10140   1/2/2022   12      Good              PM2.5   
2            10140   1/3/2022   18      Good              PM2.5   
3            10140   1/4/2022   19      Good              PM2.5   
4            10140   1/5/2022   17      Good              PM2.5   
...            ...        ...  ...       ...                ...   
1048570      13900  6/11/2015   66  Moderate              PM2.5   
1048571      13900  6/12/2015   53  Moderate              PM2.5   
1048572      13900  6/13/2015   58  Moderate              Ozone   
1048573      13900  6/14/2015   61  Moderate              Ozone   
1048574      13900  6/15/2015   34      Good              Ozone   

         Number of Sites Reporting city_ascii state_id    state_name      lat  \
0                                2   Aberdeen       WA    Washington  46.9757   
1                                2   Aberdeen       WA    Washington  46.9757   
2                                2   Aberdeen       WA    Washington  46.9757   
3                                2   Aberdeen       WA    Washington  46.9757   
4                                2   Aberdeen       WA    Washington  46.9757   
...                            ...        ...      ...           ...      ...   
1048570                          2   Bismarck       ND  North Dakota  46.8143   
1048571                          2   Bismarck       ND  North Dakota  46.8143   
1048572                          2   Bismarck       ND  North Dakota  46.8143   
1048573                          2   Bismarck       ND  North Dakota  46.8143   
1048574                          2   Bismarck       ND  North Dakota  46.8143   

              lng  population  density             timezone Month Day  Year  
0       -123.8094       16571      588  America/Los_Angeles     1   1  2022  
1       -123.8094       16571      588  America/Los_Angeles     1   2  2022  
2       -123.8094       16571      588  America/Los_Angeles     1   3  2022  
3       -123.8094       16571      588  America/Los_Angeles     1   4  2022  
4       -123.8094       16571      588  America/Los_Angeles     1   5  2022  
...           ...         ...      ...                  ...   ...  ..   ...  
1048570 -100.7694       90420      817      America/Chicago     6  11  2015  
1048571 -100.7694       90420      817      America/Chicago     6  12  2015  
1048572 -100.7694       90420      817      America/Chicago     6  13  2015  
1048573 -100.7694       90420      817      America/Chicago     6  14  2015  
1048574 -100.7694       90420      817      America/Chicago     6  15  2015  

[1048575 rows x 17 columns]>


# Remove original date column
air_quality_clean = air_quality.drop(['Date'], axis=1)
air_quality_clean


# Convert Month, Day, Year columns to number
air_quality_clean['Month'] = pd.to_numeric(air_quality_clean['Month'])
air_quality_clean['Day'] = pd.to_numeric(air_quality_clean['Day'])
air_quality_clean['Year'] = pd.to_numeric(air_quality_clean['Year'])
air_quality_clean.dtypes

CBSA Code                      int64
AQI                            int64
Category                      object
Defining Parameter            object
Number of Sites Reporting      int64
city_ascii                    object
state_id                      object
state_name                    object
lat                          float64
lng                          float64
population                     int64
density                        int64
timezone                      object
Month                          int64
Day                            int64
Year                           int64
dtype: object


# Filter to 2019 and after due to Computational resource constraint
index = np.where(air_quality_clean["Year"] >= 2019)
air_quality_clean1 = air_quality_clean.loc[index]
air_quality_clean1


# Hot Encode categorical variables 
air_quality_clean2 = pd.get_dummies(air_quality_clean1)
air_quality_clean2.dtypes

CBSA Code                         int64
AQI                               int64
Number of Sites Reporting         int64
lat                             float64
lng                             float64
                                 ...   
timezone_America/Matamoros        uint8
timezone_America/New_York         uint8
timezone_America/Phoenix          uint8
timezone_America/Puerto_Rico      uint8
timezone_Pacific/Honolulu         uint8
Length: 610, dtype: object


# Split into Train and Test
from sklearn.model_selection import train_test_split
train_data, test_data = train_test_split(air_quality_clean2, test_size=0.2)
train_data_y = train_data['AQI']
train_data_x = train_data.drop(columns=['AQI'])
test_data_y = test_data['AQI']
test_data_x = test_data.drop(columns=['AQI'])


# Convert data to Numpy array from panda dataframe
train_data_x  = np.array(train_data_x)
train_data_y  = np.array(train_data_y)
test_data_x   =  np.array(test_data_x)
test_data_y   = np.array(test_data_y)


# Normalize data to min/max scaler
from sklearn.preprocessing import MinMaxScaler
scaler=MinMaxScaler()

train_data_y = train_data_y.reshape(-1, 1)
train_data_y_s = scaler.fit_transform(train_data_y)

test_data_y = test_data_y.reshape(-1, 1)
test_data_y_s = scaler.transform(test_data_y)

train_data_x_s = scaler.fit_transform(train_data_x)
test_data_x_s = scaler.transform(test_data_x)


# Import Tensorflow libraries
import tensorflow as tf
from tensorflow import keras
from tensorflow.keras.models import Sequential
from tensorflow.keras.layers import Activation, Dense
from tensorflow.keras.optimizers import Adam
from tensorflow.keras.metrics import categorical_crossentropy


# Create Model with 3 dense layers
tf.random.set_seed(42)
air_model = tf.keras.Sequential([
    tf.keras.layers.Dense(10),
    tf.keras.layers.Dense(10),
    tf.keras.layers.Dense(1)
])


# Compile Model
air_model.compile(loss=tf.keras.losses.mae, optimizer=tf.keras.optimizers.Adam(),metrics=["mae"])


# Train Model using validation set of 0.1 to prevent over fitting
air_model.fit(x=train_data_x_s, y=train_data_y_s, validation_split=0.1, batch_size=10,epochs=10,shuffle=True,verbose=2)

Epoch 1/10
37931/37931 - 35s - loss: 0.0077 - mae: 0.0077 - val_loss: 0.0070 - val_mae: 0.0070 - 35s/epoch - 923us/step
Epoch 2/10
37931/37931 - 41s - loss: 0.0069 - mae: 0.0069 - val_loss: 0.0073 - val_mae: 0.0073 - 41s/epoch - 1ms/step
Epoch 3/10
37931/37931 - 36s - loss: 0.0069 - mae: 0.0069 - val_loss: 0.0067 - val_mae: 0.0067 - 36s/epoch - 949us/step
Epoch 4/10
37931/37931 - 36s - loss: 0.0068 - mae: 0.0068 - val_loss: 0.0067 - val_mae: 0.0067 - 36s/epoch - 954us/step
Epoch 5/10
37931/37931 - 38s - loss: 0.0068 - mae: 0.0068 - val_loss: 0.0068 - val_mae: 0.0068 - 38s/epoch - 990us/step
Epoch 6/10
37931/37931 - 42s - loss: 0.0068 - mae: 0.0068 - val_loss: 0.0067 - val_mae: 0.0067 - 42s/epoch - 1ms/step
Epoch 7/10
37931/37931 - 38s - loss: 0.0068 - mae: 0.0068 - val_loss: 0.0067 - val_mae: 0.0067 - 38s/epoch - 1ms/step
Epoch 8/10
37931/37931 - 43s - loss: 0.0068 - mae: 0.0068 - val_loss: 0.0072 - val_mae: 0.0072 - 43s/epoch - 1ms/step
Epoch 9/10
37931/37931 - 40s - loss: 0.0068 - mae: 0.0068 - val_loss: 0.0067 - val_mae: 0.0067 - 40s/epoch - 1ms/step
Epoch 10/10
37931/37931 - 36s - loss: 0.0068 - mae: 0.0068 - val_loss: 0.0067 - val_mae: 0.0067 - 36s/epoch - 944us/step

<keras.callbacks.History at 0x11fa514adc0>


# Predict on test x
predictions = air_model.predict(x=test_data_x_s,batch_size=10,verbose=2)
predictions.shape

10537/10537 - 6s - 6s/epoch - 604us/step

(105362, 1)


#Check Shape
test_data_y_s.shape

(105362, 1)


# Calculated MAE of around 0.0067, this validates the MAE we see in the model 
from sklearn.metrics import mean_absolute_error
mean_absolute_error(test_data_y_s, predictions)

0.006721152943301766

	CBSA Code	Number of Sites Reporting	lat	lng	population	density	Month	Day	Year	Category_Good	...	timezone_America/Detroit	timezone_America/Indiana/Indianapolis	timezone_America/Indiana/Vincennes	timezone_America/Juneau	timezone_America/Los_Angeles	timezone_America/Matamoros	timezone_America/New_York	timezone_America/Phoenix	timezone_America/Puerto_Rico	timezone_Pacific/Honolulu
244428	22280	2	39.5627	-119.1906	20616	65	11	7	2020	1	...	0	0	0	0	1	0	0	0	0	0
111377	30140	1	40.3412	-76.4228	80620	2388	9	18	2021	1	...	0	0	0	0	0	0	1	0	0	0
443111	29740	9	32.3265	-106.7893	129538	516	4	10	2019	0	...	0	0	0	0	0	0	0	0	0	0
279863	30780	3	34.7256	-92.3577	446814	637	4	13	2020	1	...	0	0	0	0	0	0	0	0	0	0
122409	33340	5	43.0642	-87.9675	1390046	2379	1	11	2021	0	...	0	0	0	0	0	0	0	0	0	0
...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...	...
221176	17140	13	39.1413	-84.5060	1684800	1501	3	17	2020	1	...	0	0	0	0	0	0	1	0	0	0
4806	17300	1	36.5692	-87.3413	183798	606	2	9	2022	1	...	0	0	0	0	0	0	0	0	0	0
385701	16820	1	38.0375	-78.4855	102164	1779	8	5	2019	1	...	0	0	0	0	0	0	1	0	0	0
511376	46020	1	39.3455	-120.1848	16666	199	7	1	2019	0	...	0	0	0	0	1	0	0	0	0	0
191079	49620	2	39.9651	-76.7315	236818	3211	3	8	2021	1	...	0	0	0	0	0	0	1	0	0	0

	CBSA Code	Date	AQI	Category	Defining Parameter	Number of Sites Reporting	city_ascii	state_id	state_name	lat	lng	population	density	timezone
0	10140	1/1/2022	21	Good	PM2.5	2	Aberdeen	WA	Washington	46.9757	-123.8094	16571	588	America/Los_Angeles
1	10140	1/2/2022	12	Good	PM2.5	2	Aberdeen	WA	Washington	46.9757	-123.8094	16571	588	America/Los_Angeles
2	10140	1/3/2022	18	Good	PM2.5	2	Aberdeen	WA	Washington	46.9757	-123.8094	16571	588	America/Los_Angeles
3	10140	1/4/2022	19	Good	PM2.5	2	Aberdeen	WA	Washington	46.9757	-123.8094	16571	588	America/Los_Angeles
4	10140	1/5/2022	17	Good	PM2.5	2	Aberdeen	WA	Washington	46.9757	-123.8094	16571	588	America/Los_Angeles