In part 1 of this tutorial, we explained the advantages of and proposed a methodology for using DCNNs for time series analysis by converting time series into gray-scale images. In part 2, we defined a Python class and various methods to perform data processing. In our third and final part, we will explain the topology of our model.Discuss how it is possible to establish its computation graph, and demonstrate how to run this graph in a Tensorflow session.


Our Proposed DCNN Topology

As explained in part 1 of this series, CNNs have only convoluted layers and pooling. DCNNs, on the other hand, include more diverse layers of the dense/fully connected layer. There are several combinations of layer connections that produce many different possible topologies. Additionally, selecting hyper-parameters can produce different models. So one challenge in solving our particular problem is determining how to establish an effective architecture since architectures have a great impact on model performance.

In this tutorial, we will use a simple DCNN model with two convolute layers, two pooling layers, and three dense layers. To avoid overfitting, we’ll apply a dropout technique that refers to ignoring units (i.e. neurons) during the training phase of a certain set of neurons that is chosen at random. For developing this prediction model, we adopt Tensorflow official website guidelines, which can be accessed here and here:

## building CNN model
def cnn_model():


# building computation graph

# the input images are 32*32, so flattened images are a size of 1024

                 x = tf.placeholder(tf.float32, [None, 1024])   

# because we have just two labels

                 y = tf.placeholder(tf.float32, [None, 2])

# the input images are 32*32        

                 input_layer_of_load_images = tf.reshape(x, [-1, 32, 32, 1])


# covulate layer 1 specifications are

                convolute_layer_1 = tf.layers.conv2d(

                 inputs = input_layer_of_load_images,

                 filters = 32,

                 kernel_size = [5, 5],

                 padding = "same",

                 activation = tf.nn.relu)


# pooling layer 1 specifications are

                 pool_layer_1 = tf.layers.max_pooling2d(


                 pool_size=[2, 2],



#  covulate layer 2 specifications are

                convolute_layer_2 = tf.layers.conv2d(



                kernel_size=[5, 5],




# pooling layer 2 specifications are

                pool_layer_2 = tf.layers.max_pooling2d(

                inputs = convolute_layer_2,

                pool_size=[2, 2],



# pooling layer 2 output is going to be flattened as follows to be the input for the next layer

                 pool_layer_2_flat = tf.reshape(pool_layer_2,

                 [-1, 8 * 8 * 64])


# dense layer 1 specifications are

                 dense_layer_1 = tf.layers.dense(inputs = pool_layer_2_flat,




# dense layer 2 specifications are

               dense_layer_2 = tf.layers.dense(inputs = dense_layer_1,

               units = 600,



# using dropout technique to avoid overfitting, which shuts down neurons at a rate of 40%

                dropout = tf.layers.dropout(

                inputs = dense_layer_2, rate = 0.4)


# dense layer 3 specifications are

                dense_layer_3 = tf.layers.dense(inputs = dropout,



# prediction and final outputs

               prediction = tf.layers.dense(dense_layer_3, activation=tf.nn.softmax, units = 2)

              cross_entropy = tf.reduce_mean(-tf.reduce_sum(y * tf.log(prediction), reduction_indices=[1]))  

              train_step = tf.train.AdamOptimizer(0.001).minimize(cross_entropy)

              number_of_epochs = 15


# creating  session for running computation graph

               with tf.Session() as sess:


                               for epoch in range(number_of_epochs):

                                       epoch_loss = 0

                                       makeImage.index_of_columns_2_be_batched = [i for i in range(0, X_train.shape[0])]

                                       for _ in range(number_of_batches):

                                                      epoch_x, epoch_y = makeImage.next_batch(X_train, y_train, batch_size)

                                                      _, c =[train_step, cross_entropy], feed_dict={x: epoch_x, y: epoch_y})

                                                     epoch_loss += c

                              print('Epoch', epoch + 1, 'completed out of', number_of_epochs,' with loss:',epoch_loss)

                              correct = tf.equal(tf.argmax(prediction, 1), tf.argmax(y, 1))

                              accuracy = tf.reduce_mean(tf.cast(correct, 'float'))


print('Accuracy in epoch', epoch + 1, 'is:',accuracy.eval({x:X_test, y:y_test}))




After running the model, the following results might be produced. (Your own results will be different than the following due to the random configuration we deployed while building the model.)

Epoch 1 completed out of 15  with loss: 51.4090359211

Accuracy in epoch 1 is: 0.853059

Epoch 2 completed out of 15  with loss: 36.055752486

Accuracy in epoch 2 is: 0.862016

Epoch 3 completed out of 15  with loss: 30.448112756

Accuracy in epoch 3 is: 0.87955

Epoch 4 completed out of 15  with loss: 24.6789700091

Accuracy in epoch 4 is: 0.885649

Epoch 5 completed out of 15  with loss: 19.7951619923

Accuracy in epoch 5 is: 0.874404

Epoch 6 completed out of 15  with loss: 15.8194133416

Accuracy in epoch 6 is: 0.883362

Epoch 7 completed out of 15  with loss: 12.7372778915

Accuracy in epoch 7 is: 0.877644

Epoch 8 completed out of 15  with loss: 10.4595579039

Accuracy in epoch 8 is: 0.881456

Epoch 9 completed out of 15  with loss: 7.60100067966

Accuracy in epoch 9 is: 0.879169

Epoch 10 completed out of 15  with loss: 6.00782689825

Accuracy in epoch 10 is: 0.877835

Epoch 11 completed out of 15  with loss: 5.18157571554

Accuracy in epoch 11 is: 0.872689

Epoch 12 completed out of 15  with loss: 4.55696826754

Accuracy in epoch 12 is: 0.883362

Epoch 13 completed out of 15  with loss: 3.64296960807

Accuracy in epoch 13 is: 0.876882

Epoch 14 completed out of 15  with loss: 2.26440465293

Accuracy in epoch 14 is: 0.883743

Epoch 15 completed out of 15  with loss: 2.38258726092

Accuracy in epoch 15 is: 0.879931


In this three-part blog, we provided a tutorial on how to use DCNNs for time series analysis. We first provided information on our proposed methodology, explaining how it’s possible to convert a time series into gray-scale images. Next, we provided Python codes for data processing and preparation. Finally, we used Tensorflow to complete our predictions. As seen in our results, this method allows us to reach an accuracy of about 88% when predicting trends for the next state of electricity consumption.

At Avenue Code, our data science teams enjoy providing solutions for real-world challenges by using DCNNs and many other technologies. We serve diverse industries, including retail, energy, telecommunications, etc., and we invite you to contact us to discover our solutions to your challenges.


Hossein Javedani Sadaei

Hossein Javedani Sadaei is a Senior Data Scientist at Avenue Code with a post-doctoral in big data mining and a PhD in statistics. He works mostly with machine learning and deep learning in retail, telecommunication, energy, and stock. His main expertise is developing scalable machine learning and deep learning algorithms using fuzzy logics.