sadhityadimas / weather-time-series-analysis

New York City Taxi Trip - Hourly Weather Data Time Series Analysis

Weather-Time-Series-Analysis

New York City Taxi Trip - Hourly Weather Data Time Series Analysis

For Dicoding Submission Project

• Dataset is provided by Kaggle uploaded by Meinertsen
• Link to File : https://www.kaggle.com/meinertsen/new-york-city-taxi-trip-hourly-weather-data
• The dataset can be fetch using kaggle library on colab
• But i also provided the extracted dataset (weather.csv)
• 10.400 samples
• 80:20 train test split

Library Used

• Tensorflow
• Keras
• Scikit learn
• matplotlib
• Numpy
• Pandas

Data set composition

• Variable used: tempm (hourly mean temperature)
• Data Shape : (10476, 2)

Step by Step Model Building

Data Cleaning

• Check Null
• Drop NUll (5 rows)
• Check Dtypes\

Examine the data

Set Minimum MAE

• (Variable maximum value - variable minimum value) * (10/100)

Perform Train-Test Split

• 0.2 validation set
• Shuffle False (time series data is sensitive to shuffled data)
• set Random state

Build LSTM model

• 2 hidden layer (perceptron 512 units and 128 units)
• output layer 'Softmax'
• model = keras.Sequential([
• layers.Conv1D(filters=32, kernel_size=5,

•                  strides=1, padding="causal",
  

•                  activation="relu",
  

•                  input_shape=[None, 1]),
  

• layers.LSTM(64, return_sequences=True),
• layers.LSTM(64, return_sequences=True),
• layers.Dense(30, activation="relu"),
• layers.Dense(10, activation="relu"),
• layers.Dense(1),
• layers.Lambda(lambda x: x * 400)
• ])

Adding Loss Function and Optimizer with learning rate

• loss : Huber
• optimizer Sgd with learning rate
• metrics MAE

Adding custom Early stop function using tensorflow callback

• Set to stop training if val accuracy is smaller than target MAE

Training the model

• epoch : 500
• callbacks

Results

• Best Validation Loss: 4.91

• Best Validation MAE: 5.37

• Loss

• Accuracy