How would one use a RNN when predicting temperature?

Let’s assume you have the following data structure and we want to predict the temperature given 1 day in the past:

import tensorflow as tf
import pandas as pd
import numpy as np

df = pd.DataFrame(data={
    'temperature': np.random.random((1, 20)).ravel(),
    'pressure': np.random.random((1, 20)).ravel(),
    'humidity': np.random.random((1, 20)).ravel(),
    'wind': np.random.random((1, 20)).ravel()
})

print(df.to_markdown())

The first thing we have to do is separate the data into features and labels:

features = df.iloc[::2, :] # Get every first row 
labels = df.iloc[1::2, :] # Get every second row since we want to predict the temperature given 1 day in the past

Features:

Labels:

Since you are only interested in predicting the temperature, we can remove the other features from the labels and convert both to arrays:

features = features.to_numpy() # shape (10, 4)
labels = labels['temperature'].to_numpy() # shape (10,)
features = np.expand_dims(features, axis=1) # shape (10, 1, 4)

Note that a time dimension is added to features, which essentially means that each sample in the dataset represents one timestep (one day) and for each timestep there are 4 features (temperature, pressure, humidity, wind).

Building and running a RNN model:

inputs = tf.keras.layers.Input(shape=(features.shape[1], features.shape[2]))
rnn_out = tf.keras.layers.SimpleRNN(32)(inputs)
outputs = tf.keras.layers.Dense(1)(rnn_out) # one output = temperature

model = tf.keras.Model(inputs=inputs, outputs=outputs)
model.compile(optimizer='adam', loss="mse")
model.summary()
history = model.fit(features, labels, batch_size=2, epochs=3)

Model: "model_1"
_________________________________________________________________
 Layer (type)                Output Shape              Param #   
=================================================================
 input_2 (InputLayer)        [(None, 1, 4)]            0         
                                                                 
 simple_rnn (SimpleRNN)      (None, 32)                1184      
                                                                 
 dense_1 (Dense)             (None, 1)                 33        
                                                                 
=================================================================
Total params: 1,217
Trainable params: 1,217
Non-trainable params: 0
_________________________________________________________________
Epoch 1/3
5/5 [==============================] - 1s 9ms/step - loss: 0.7859
Epoch 2/3
5/5 [==============================] - 0s 7ms/step - loss: 0.5862
Epoch 3/3
5/5 [==============================] - 0s 6ms/step - loss: 0.4354

Make predictions like this:

samples = 1
model.predict(tf.random.normal((samples, 1, 4)))
# array([[-1.610171]], dtype=float32)

You can also consider normalizing your data before training like this:

# You usually also normalize your data before training
mean = df.mean(axis=0)
std = df.std(axis=0)
df = df - mean / std

And that’s about it.