In [1]:

# Setup logging
from timeseria import logger
logger.setup(level='INFO')

# Set default plot type as image
import os
os.environ["DEFAULT_PLOT_TYPE"] = "image"

⚠️ In this notebook, plots are configured to be rendered as images because otherwise they won't display correctly on GitHub or when the notebook is loaded (usually because not trusted). To get interactive plots, remove the line above (or change it to "interactive") and re-run the notebook.

Temperature and humidity forecasting with LSTM¶

This notebook showcases the LSTM forecatser of Timeseria.

Let's start by loading an example time series, and resample it to one hour:

In [2]:

from timeseria import TEST_DATASETS_PATH
from timeseria.datastructures import TimeSeries
timeseries = TimeSeries.from_csv(TEST_DATASETS_PATH + 'humitemp_long.csv').resample('1h')

[INFO] timeseria.transformations: Using auto-detected sampling interval: 615.0s
[INFO] timeseria.transformations: Resampled 17910 DataTimePoints in 3191 DataTimePoints

Have a look at the time series

In [3]:

timeseries

Out[3]:

Time series of #3191 points at 1h resolution, from point @ 1546477200.0 (2019-01-03 01:00:00+00:00) to point @ 1557961200.0 (2019-05-15 23:00:00+00:00)

Plot the time series

In [4]:

timeseries.plot()

Out[4]:

Model fit, usage and evaluation¶

Instantiate and fit the model:

In [5]:

from timeseria.models import LSTMForecaster
forecaster = LSTMForecaster(window=24, neurons=64, features=['values', 'hours'])
forecaster.fit(timeseries, epochs=30, verbose=True, reproducible=True)

Epoch 1/30
93/93 [==============================] - 1s 4ms/step - loss: 0.0198
Epoch 2/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0039
Epoch 3/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0034
Epoch 4/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0031
Epoch 5/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0030
Epoch 6/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0028
Epoch 7/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0026
Epoch 8/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0026
Epoch 9/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0024
Epoch 10/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0022
Epoch 11/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0022
Epoch 12/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0021
Epoch 13/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0020
Epoch 14/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0019
Epoch 15/30
93/93 [==============================] - 0s 5ms/step - loss: 0.0018
Epoch 16/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0018
Epoch 17/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0017
Epoch 18/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0017
Epoch 19/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0017
Epoch 20/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0017
Epoch 21/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0016
Epoch 22/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0016
Epoch 23/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0015
Epoch 24/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0015
Epoch 25/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0015
Epoch 26/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0014
Epoch 27/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0014
Epoch 28/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0014
Epoch 29/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0014
Epoch 30/30
93/93 [==============================] - 0s 4ms/step - loss: 0.0014

Call the predict() function of the model. This returns key-value data with the prediction values

In [6]:

forecaster.predict(timeseries)

Out[6]:

{'humidity': 41.45781251050483, 'temperature': 22.01087811118259}

The timeseries on which to call the predict has to be at least of 24 hours (the window length)

In [7]:

try:
    forecaster.predict(timeseries[0:23])
except Exception as e:
    print(e)

The series length (23) is shorter than the model window (24)

The forecats() method returns not only data but also their data points, with the timestamp

In [8]:

forecaster.forecast(timeseries)

Out[8]:

Time point @ 1557964800.0 (2019-05-16 00:00:00+00:00) with data "{'humidity': 41.45781251050483, 'temperature': 22.01087811118259}"

The apply() method allows instead to directly apply the model on a time series (by combining the forecasts with the real observations).

In [9]:

forecaster.apply(timeseries, steps=12)[-1000:].plot()

Out[9]:

Evaluate the forecaster. By default, forecasters get evaluated with RMSE and MAE error metrics:

In [10]:

forecaster.evaluate(timeseries)

Out[10]:

{'humidity_RMSE': 1.15405250053555,
 'humidity_MAE': 0.8408280784195243,
 'temperature_RMSE': 0.3158130676246483,
 'temperature_MAE': 0.22328342716990682}

For more advanced modes of evaluation, including using other error metrics, plotting the error distribution and/or the predictions, check out the "Forecasting - Advanced Evaluation" notebook.