I'm developing a python program which predict daily timeseries values.
Each daily timeseries contains 288 values (a record every 5 minutes).

The main idea is to train a LSTM model with 7 days data and predict the next day. After the prediction, train again the model with the last 7 days (so the one just predicted and discarding the first one).

In order to do this, I have implemented the following code:

for idx in range(len(self.test)):
    graph = tf.Graph()
    with tf.Session(graph=graph):

        history = [x for x in self.train]

        # prepare data
        train_x, train_y = to_supervised(self.train, self.n_input, self.n_output)
        self.fit_model(train_x, train_y)

        # make the prediction
        yhat_sequence = self.forecast(history, self.n_input)

        real_sequence = self.test[idx, :].reshape(-1, len(self.test[idx, :]))
        real_sequence = self.scaler_test.inverse_transform(real_sequence)

        yhat_sequence = yhat_sequence.reshape(-1, len(yhat_sequence))
        yhat_sequence = self.scaler_train.inverse_transform(yhat_sequence)

        # add real sequence to history
        history.append(self.test[idx, :])

        # add real sequence to train
        last_day_scaled = np.array([self.test[idx, :]])

        self.train = np.concatenate((self.train, last_day_scaled))

        # remove first day from train in order to have always one week of training
        self.train = self.train[1:, :]

        # reset model in order to train it again
        self.model = None

Here the fit_model function:

def fit_model(self, train_x, train_y):
    # define parameters
    verbose, epochs, batch_size = 1, 100, 128

    # reshape trains 

    n_timesteps, n_features, n_outputs = train_x.shape[1], train_x.shape[2], train_y.shape[1]

    if self.model is None:
        print("Create model")

        self.model = Sequential()
        self.model.add(LSTM(100, input_shape=(n_timesteps, n_features)))
        self.model.add(LSTM(100, return_sequences=True))
        self.model.compile(loss='mse', optimizer='adam')

    # fit network
    self.model.fit(train_x, train_y, epochs=epochs, batch_size=batch_size, verbose=verbose)

This codes works without errors, but predictions done training after training are very similar to each other even if training data change.

Pictures below are three different days (blu real data and green prediction) enter image description here enter image description here enter image description here

I've try to increase epochs, neurons, add DroupsOut and to increase number of training days

Is there something wrong with my code? How can I improve my prediction training after training?

  • $\begingroup$ Can you show some of your predictions $\endgroup$ – Sachin Yadav Nov 4 '19 at 17:05
  • $\begingroup$ yep, i've added three days of prediction $\endgroup$ – Giordano Nov 4 '19 at 17:49

Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Browse other questions tagged or ask your own question.