Temperature TimeSeries Prediction

Author: Bertrand Delorme

Date: 23/05/2018

General

Goal of the project

Predicting the temperature of a subway station at time h+1 based on historical data from the beginning of time to time h.

Dataset

The dataset is provided by RATP (link) under open data Etalab license and contains air property measurements from a subway station in csv format. It has 7 features: day and time of measurement, temperature, humidity, particulates concentration and concentration of 3 chemical tracers (NO, NO2, CO2).

Implementation choices

Two implementations have been tested to predict the temperature at h+1:

• a random forest ensemble algorithm based on the value of temperature at h and h-1, the hour of the day, and the day of the month.
• a Seasonal AutoRegressive Integrated Moving Average (SARIMA) algorithm.

As the random forest algorithm performs better in our experiments, we use it as finale response for the coding challenge.

Installation

Requirements

The dependencies needed are:

• numpy
• scipy
• pandas
• scikit-learn
• jupyter
• statsmodels

as shown in config.yml.

Setup instructions

We suggest using Anaconda to create a conda environment with the required dependencies as:

cd Temperature_TS_Prediction
conda env create -f config.yml
source activate Temperature_TS_Prediction

Project Overview

Libraries used

• numpy for efficient data structures and functions for scientific computing.
• pandas for data manipulation and data analysis.
• matplotlib for data visualization.
• scikit-learn for Random Forest implementation.
• statsmodels for Seasonal Arima implementation.

Architecture

• data/: contains the raw historical dataset.
• src/: contains the source code to make the prediction.
• config.yml: conda config file.
• predict: executable to get the prediction.
• exploration.ipynb: jupyter notebook showing the exploration process and justification of implementation choices.

How it works

Given a dataset with historical data until time h, you get a prediction for time h+1 by simply doing:

./predict path_to_historical_dataset

This should last less than a minute and print the prediction at time h+1. To store the result in a file result.txt, do:

./predict path_to_historical_dataset > result.txt

We suggest that you first go over the jupyter notebook to understand the choices made.

Improvements to do

• test files
• Random Forest in Fourier Space
• implementation of RNN with LSTM