ml

machine learning for time series Prediction

This project shows how to do time series data prediction with machine learning models. The data is Historical Hourly Weather Data 2012-2017 from Kaggle.

Two regression models of Tensorflow are used in this project:

• DNNRegressor
• LinearRegressor

These two models are used to do one step ahead prediction.

Prediction performance

Three models are used to evaluate the performance:

• The Baseline method is to use the value of last step as the predicted value of next step;
• The DNN model is a simple shallow nerual networks with three hidden layers: 32 x 16 x 16;
• The Linear model has the same feature set as the DNN model.

Following table gives the inital RMSE of the three models without much tune:

method	train-RMSE	test-RMSE
Baseline	-	1.6046
DNN	1.3849	1.1665
Linear	1.3946	1.1378
Holt-Winter-D	1.9	1.6045
Holt-Winter	1.9	1.6063

Figure-1 DNN model for Hourly temperature prediction.

Figure-2 Holt-Winter-D model for Hourly temperature prediction.

Train the models

0. Download the dataset

Download temperature.csv from Kaggle project; and put it to ./data/temperature.csv

1. Extract the data for one city

python extract.py

By default, it will extract the data for Denver, output=./data/denver.csv

2. Generate training/test data sets

python gen_data.py

This will genereate features for the training data and test data. Currently, the features are:

• the values of last 10 time steps;
• the deltas of the the last 10 time steps;
• the 10 values of previous days aroud current time;
• the deltas the previous days values;
• the month of the data;
• the day of the data;
• the hour of the data;

By default, output will be ./data/denver-features-test.csv and ./data/denver-features-train.csv

3. Train the model

python main.py

DNNRegressor or LinearRegressor can be selected in main.py manually before the training.

4. calculate the RMSE and draw the data

python compare.py