LSTMs in time series forecasting

In the NLP sections, we discussed the capabilities of LSTMs and their improvements over RNNs by mitigating issues such as the vanishing gradient problem, enabling the model to learn longer sequences. In the context of time series forecasting, LSTM networks can be quite powerful. Let’s see how we can apply LSTMs to our sales dataset:

1. Let’s begin by preparing our data:

   # Create sequences

   seq_len = 20

   X = []

   y = []

   for i in range(seq_len, len(data)):

       X.append(data[i-seq_len:i])

       y.append(data[i])

   X = np.array(X)

   X = X.reshape(X.shape[0], X.shape[1], 1)

   y = np.array(y)

   # Train/val split

   split = int(0.8*len(X))

   X_train, X_val = X[:split], X[split:]

   y_train, y_val = y[:split], y[split:]

   # Set batch size and buffer size

   batch_size = 64

   buffer_size = 1000

   # Create dataset

   dataset = tf.data.Dataset.from_tensor_slices(

       (X_train, y_train))

   dataset = dataset.shuffle(

       ...