50
Internet of Things (IoT)
filter over the dataset used. The filter is thought of as an array of 2-D data of the size 3 × 3,
5 × 5, 7 × 7, etc. The set of nodes obtained from the convolution process is termed as feature
maps. While the dataset used is the input to the first hidden layer, the number of feature
maps obtained from the previous layers becomes the input for the subsequent hidden layers. At the same time, for each layer, the number of output feature maps shall be the same
as the number of filters that is to be used in the convolution process. Based on the type of
convolution function used in the process, the size of the output feature maps can be determined. As the feature extraction process depends on the size and number of filters used in
each hidden layer, processing the large dataset with huge parameters demand a high-end
computing environment with large memory capabilities for providing optimum performance (Panda, 2016). In this, use of the random filter is a good choice that is represented in
the range of [−1 to +1]. The output from the convolution process is applied to an activation
function for linear or nonlinear transformation across different layers so that the output
falls in the range of [−1,+1] or [0,1]. Although several activation functions are available, we
use rectified linear unit activation function for our analysis which says that for input x and
output y, f(y) = 1, if x>0, and f(y) = 0, otherwise. Finally, pooling function is used (here, we
use max-pooling) for down sample and to downsize the input features’ map to half of its
original size. Dimensionality reduction on the datasets shall be achieved by the end of this
stage.
The fully connected layer is the last but one to the DCNN layers where the result is
obtained as a single vector after consolidating the previous layers single-node output feature maps. As this stage leads to classification of the dataset, proper weights and bias
are applied to this layer so that the cost of misclassification falls below a certain threshold or else the back-propagation process gets initiated for better weight and bias updates
across layers for error minimization. Finally, the output layer presents the desired output. Different hyperparameters such as learning rate and momentum are then specified to
have a faster convergence.
3.6.4 Smart Grid
Cloud computing also has many useful applications in smart grid, and data mining is one
of its most significant attributes. A smart grid generates huge volumes of data through the
weather conditions, solar or wind characteristics, network intrusions, people’s electricity
consumption, and how much electrical power people add from their own rooftop systems
continuously to the grid. Utilities will need to store that data, and cloud computing is the
solution for that. Once smart grid data are stored in the cloud, utilities will need data mining techniques to develop knowledge from the raw data. They will use the data to determine the relationship between demand and supply or to explain customer impressions
or opinions about their smart grid services. So cloud computing, data mining, and smart
grids are all very closely related.
It is worth noting that the quality of the intelligence we gain from data mining will be
influenced by the quality and quantity of data we have available to us. Smart grid is creating a wonderful database that is a resource for monitoring the system and even solving
operational problems. But the database needs to be comprehensive and demands a good
database for every section of the smart grid. This way, it is envisaged that “Internet of
Things” will help implement smart grids
The IoT generally connects to any object, whether the object is an element in the smart
grid, a human being, or any other physical entity. It is capable enough to transfer the data
Internet of Things (IoT)
filter over the dataset used. The filter is thought of as an array of 2-D data of the size 3 × 3,
5 × 5, 7 × 7, etc. The set of nodes obtained from the convolution process is termed as feature
maps. While the dataset used is the input to the first hidden layer, the number of feature
maps obtained from the previous layers becomes the input for the subsequent hidden layers. At the same time, for each layer, the number of output feature maps shall be the same
as the number of filters that is to be used in the convolution process. Based on the type of
convolution function used in the process, the size of the output feature maps can be determined. As the feature extraction process depends on the size and number of filters used in
each hidden layer, processing the large dataset with huge parameters demand a high-end
computing environment with large memory capabilities for providing optimum performance (Panda, 2016). In this, use of the random filter is a good choice that is represented in
the range of [−1 to +1]. The output from the convolution process is applied to an activation
function for linear or nonlinear transformation across different layers so that the output
falls in the range of [−1,+1] or [0,1]. Although several activation functions are available, we
use rectified linear unit activation function for our analysis which says that for input x and
output y, f(y) = 1, if x>0, and f(y) = 0, otherwise. Finally, pooling function is used (here, we
use max-pooling) for down sample and to downsize the input features’ map to half of its
original size. Dimensionality reduction on the datasets shall be achieved by the end of this
stage.
The fully connected layer is the last but one to the DCNN layers where the result is
obtained as a single vector after consolidating the previous layers single-node output feature maps. As this stage leads to classification of the dataset, proper weights and bias
are applied to this layer so that the cost of misclassification falls below a certain threshold or else the back-propagation process gets initiated for better weight and bias updates
across layers for error minimization. Finally, the output layer presents the desired output. Different hyperparameters such as learning rate and momentum are then specified to
have a faster convergence.
3.6.4 Smart Grid
Cloud computing also has many useful applications in smart grid, and data mining is one
of its most significant attributes. A smart grid generates huge volumes of data through the
weather conditions, solar or wind characteristics, network intrusions, people’s electricity
consumption, and how much electrical power people add from their own rooftop systems
continuously to the grid. Utilities will need to store that data, and cloud computing is the
solution for that. Once smart grid data are stored in the cloud, utilities will need data mining techniques to develop knowledge from the raw data. They will use the data to determine the relationship between demand and supply or to explain customer impressions
or opinions about their smart grid services. So cloud computing, data mining, and smart
grids are all very closely related.
It is worth noting that the quality of the intelligence we gain from data mining will be
influenced by the quality and quantity of data we have available to us. Smart grid is creating a wonderful database that is a resource for monitoring the system and even solving
operational problems. But the database needs to be comprehensive and demands a good
database for every section of the smart grid. This way, it is envisaged that “Internet of
Things” will help implement smart grids
The IoT generally connects to any object, whether the object is an element in the smart
grid, a human being, or any other physical entity. It is capable enough to transfer the data