0

I am working on a neural network model to predict natural gas future price after 2020 using pre-pandemic data. The results seems impressive. But the learning process seems questionable. the error usually stopped to improve since 20 runs(epoch) even I set epochs = 2000. Change to adam does not help either [loss history][test results]

NN_model = Sequential()
NN_model.add(Dense(32, kernel_initializer='normal',input_dim = train_data.shape2,activation='elu'))
NN_model.add(Dense(128, kernel_initializer='normal',activation='elu'))
NN_model.add(Dense(64, kernel_initializer='normal',activation='elu'))
NN_model.add(Dense(16, kernel_initializer='normal',activation='elu'))
NN_model.add(Dense(1, kernel_initializer='normal',activation='linear'))
NN_model.compile(loss='mean_absolute_error', optimizer='rmsprop', metrics=['mean_absolute_error'])
NN_model.summary()

checkpoint_name = 'Weights-{epoch:03d}--{val_loss:.5f}.hdf5'
checkpoint_path = os.path.join('models/', checkpoint_name)
checkpoint = ModelCheckpoint(checkpoint_path, monitor='val_loss', verbose = 1, save_best_only = True, mode ='auto')
callbacks_list = [checkpoint]

history = NN_model.fit(train_data, target, epochs=2000, batch_size=64, validation_split = 0.2, callbacks=callbacks_list).history
plot_loss(history)

eddilicious
  • 1
  • sorry, I forgot to ask the question. Can someone point out what kind of settings in my model can be improved so it can continue reduce its loss into at least epoch 200 - 500s? – eddilicious Jan 10 '22 at 05:14
  • Have you tried training it for 20k epochs instead? You will find more hints in the linked thread. – Tim Jan 10 '22 at 07:48
  • I actually found the issue, the data was not balanced. increase the batch size provide band-it, balance the data improve it in every run. did not change the learn rate. – eddilicious Jan 11 '22 at 06:13

0 Answers0