Daily Minimum Temperatures Forecasting

Forecasting is such an essential part of business and everyday life. It is in your weather app and your business plan. I did some research and discovered that Facebook open sourced their time series forecasting model Prophet. Read More

Getting Started

• Python 3.7 - More info
• prophet installation guide - More Info

Libraries to install

pip install -r requirements.txt

or pip install

pandas
glob2
numpy
fbprophet
plotly

You might find that fbprohet does not install via pip then you can run a conda-forge:

conda install -c conda-forge fbprophet

Download the data

• Time Series Minimum Temp 1981-1991

The Model

You can find the code in the model model.py file

Import the libraries and data

import pandas as pd
import glob2 as glob
import numpy as np
from fbprophet import Prophet
from fbprophet.plot import add_changepoints_to_plot
from fbprophet.plot import plot_plotly
import plotly.express as px
import plotly.offline as py

The data requires some very basic preprocessing, including formatting and dtype changes.

#####################################################################################################################################
### IMPORT DATA ###
#####################################################################################################################################
df = pd.read_csv('temp_data.csv')

#convert string to date format
df['Date'] = pd.to_datetime(df['Date'])
df['Daily minimum temperatures'] = df['Daily minimum temperatures'].map(lambda x: x.lstrip('?'))
df['Daily minimum temperatures'] = df['Daily minimum temperatures'].astype(float)

#####################################################################################################################################
### REVIEW THE TIMESERIES DATA ###
#####################################################################################################################################

#plot the daily minimum temperatures
fig = px.line(df, x='Date', y='Daily minimum temperatures')
fig.show()

Fit and Review Model

Prophet requires a specific format of data and column names to passed, in order to fit. The model is set to default for the first round of assessment. When working with timeseries data it is important to understand trend and seasonality. This can be broken down by decomposing the timeseries dataset.

#####################################################################################################################################
### Fit the Model ###
#####################################################################################################################################

#format the data to required format for the library {Date: 'DS', Variable: 'Y'}
df = df.rename(columns={'Date':'ds', 'Daily minimum temperatures': 'y'})
forecast_training = df[['ds','y']]
forecast_training['y'] = np.where(
    forecast_training['y'] == 0, 0.1, forecast_training['y']
) #remove 0 values - otherwise it isn't possible to cross validate the model

#fit model
m = Prophet()
m.fit(forecast_training)

#####################################################################################################################################
### Prediction incl. Parameters ###
#####################################################################################################################################

future = m.make_future_dataframe(periods=365)
future.tail() #tail end of the data

forecast = m.predict(future)
forecast[['ds', 'yhat', 'yhat_lower', 'yhat_upper']].tail()

#####################################################################################################################################
### REVIEW MODEL PERFORMANCE ###
#####################################################################################################################################

#plot the forecast
fig1 = m.plot(forecast)

Temperatures obviously have a yearly seasonality linked to the seasons (winter,summer,etc.). The specific dataset is from Melbourne (Australia), which is in the Southern hemisphere. Meaning that it will be colder between April and August. This is displayed in the yearly trend breakdown in the figure below. This means we should have 180 horizons in the data.

#plot decomposition of seasonality and trends
fig2 = m.plot_components(forecast)

#Let's take a closer look
py.init_notebook_mode()

fig = plot_plotly(m, forecast)  # This returns a plotly Figure - interactive plot - use the date slider at the bottom
py.iplot(fig)

Run diagnostics on the first model

#####################################################################################################################################
### RUN DIAGNOSTICS ###
#####################################################################################################################################

#run cross validation
from fbprophet.diagnostics import cross_validation
df_cv = cross_validation(m, initial='720 days', period='180 days', horizon = '180 days')
df_cv.head()

#plot cross validation absolute percent error for forecast
from fbprophet.plot import plot_cross_validation_metric
fig = plot_cross_validation_metric(df_cv, metric='mape')

Adjust seasonality and re-forecast

More info on Fourier

#####################################################################################################################################
### ADJUST & RE-FORECAST ###
#####################################################################################################################################

#Fourier Order for Seasonalities
from fbprophet.plot import plot_yearly
m = Prophet(yearly_seasonality=20).fit(forecast_training)
a = plot_yearly(m)

#Specifying Custom Seasonalities - Update forecast parameters
m = Prophet(weekly_seasonality=False)
m.add_seasonality(name='yearly', period=365, fourier_order=8)
forecast = m.fit(forecast_training).predict(future)
fig = m.plot_components(forecast)

Final Forecast DataFrame

final_df = forecast[['ds','yhat','yhat_lower','yhat_upper']].rename(
    columns={'ds':'Date','yhat':'Daily minimum temperatures'}
)

#plot the daily minimum temperatures - Interactive plot
import plotly.graph_objects as go

fig = go.Figure()
fig.add_trace(go.Scatter(x=final_df.Date, y=final_df['yhat_upper'], name="Upper",
                         line_color='deepskyblue'))

fig.add_trace(go.Scatter(x=final_df.Date, y=final_df['yhat_lower'], name="Lower",
                         line_color='dimgray'))

fig.add_trace(go.Scatter(x=final_df.Date, y=final_df['Daily minimum temperatures'], name="Daily minimum temperatures",
                         line_color='black'))

fig.update_layout(title_text='Daily minimum temperatures',
                  xaxis_rangeslider_visible=True)
fig.show()

Author

• Francois van Heerden - Experience - LinkedIn Profile

Acknowledgments

• Found inspiration from multiple fellow Data Scientists in the open source community