**5. Mobility in wireless sensor networks**

Mobility may be an important consideration in the energy conservation scheme of sensor networks. Mobility becomes necessary in energy conservation when the

**249**

*Applications of Prediction Approaches in Wireless Sensor Networks*

sink or nodes are mobile in the network [28, 29]. Static nodes with multi-hop communication usually have the energy hole problem where nodes closer to the sink are depleted of their energy [30]. The energy hole problem occurs when nodes further away from the sink use them as intermediate nodes to hop data to the sink. Mobile sinks, therefore, go round the network to collect data samples from the nodes to reduce the cost of communication and hence energy lost [31]. The challenge of implementing mobile sinks is to ensure optimum paths are maintained within the

The design, deployment and evaluation of sensor nodes in WSNs depend on the environment and the designed objectives for the intended application (e.g., water quality monitoring, fire detection applications, etc.). The sensor network in their deployable environments becomes successful when the stakeholders take into account the network size, the topology, and the communication models used for achieving the application goals. Mobile sensor nodes move to form mobile sensor networks and reposition/reorganize themselves in the sensor network. For example, nodes placed in the ocean are capable of measuring parameters such as ocean current speed, temperature, salinity, pressure, and other chemicals. These sensors move about collecting data in real-time and transmitting the data to a central

Terrestrial WSNs mobility occurs differently from aquatic WSN mobility. For example, sensor nodes deployed in freshwater sources are affected by the water current. The node movement affects the sensor network design, the distance between the nodes (i.e., during route discovery), propagation, energy utilization among others. In an aquatic environment, electromagnetic wave propagation to sinks above water is challenging. Energy conservation and mobility in this type of network create challenges in the design of routing protocols. Routing protocols that are capable of minimizing the energy consumed by nodes, managing the random variation in the network topology and minimizing the delay in communication are therefore

In designing and evaluating mobility models for aquatic environments, our work took into consideration the conditions and characteristics of freshwater sources. Freshwater sources such as rivers are in constant motion. The velocity of the river depends on the slope of the land, the size and shape of the bed, and the quantity of water in the river [4, 32]. Sensor nodes may be deployed in one of the following environments; 1) slow but deep water bodies, 2) swift but deep water bodies, and 3) swift but shallow water bodies. Freshwater sources are characterized by three key factors: velocity, gradient, and discharge. Velocity is the distance that the water in a river travels in a given amount of time. The velocity relates to the energy levels of the water in the river. For example, objects (small or large) deposited in a swift or fast-moving river are carried downstream or along the river path quickly as compared to slow-moving rivers. The velocity of a river is affected by the gradient, discharge, and the shape of the river path that the water travels. The gradient is a measure of the steepness of a river, and a river's discharge is expressed as the quantity of water that moves through the different points along the river path at any

given time. The discharge varies along the length of the river [32].

Mobile sensor nodes continue to move with water currents after the initial deployment. There are different mobility models used to simulate WSN projects. These include constant acceleration, constant position, constant velocity, Gauss Markov, hierarchical, random direction 2D, random waypoint, steady state random waypoint, random walk2D, and waypoint mobility models. Although these mobility

*DOI: http://dx.doi.org/10.5772/intechopen.94500*

repository for real-time data analysis [28].

required in an aquatic environment.

network [29].

**5.1 Mobility models**

sink or nodes are mobile in the network [28, 29]. Static nodes with multi-hop communication usually have the energy hole problem where nodes closer to the sink are depleted of their energy [30]. The energy hole problem occurs when nodes further away from the sink use them as intermediate nodes to hop data to the sink. Mobile sinks, therefore, go round the network to collect data samples from the nodes to reduce the cost of communication and hence energy lost [31]. The challenge of implementing mobile sinks is to ensure optimum paths are maintained within the network [29].

## **5.1 Mobility models**

*Wireless Sensor Networks - Design, Deployment and Applications*

**248**

**Figure 6.**

**Figure 5.**

*WSN deployment techniques.*

**5. Mobility in wireless sensor networks**

*Energy consumption by the components of a sensor node.*

Mobility may be an important consideration in the energy conservation scheme of sensor networks. Mobility becomes necessary in energy conservation when the

The design, deployment and evaluation of sensor nodes in WSNs depend on the environment and the designed objectives for the intended application (e.g., water quality monitoring, fire detection applications, etc.). The sensor network in their deployable environments becomes successful when the stakeholders take into account the network size, the topology, and the communication models used for achieving the application goals. Mobile sensor nodes move to form mobile sensor networks and reposition/reorganize themselves in the sensor network. For example, nodes placed in the ocean are capable of measuring parameters such as ocean current speed, temperature, salinity, pressure, and other chemicals. These sensors move about collecting data in real-time and transmitting the data to a central repository for real-time data analysis [28].

Terrestrial WSNs mobility occurs differently from aquatic WSN mobility. For example, sensor nodes deployed in freshwater sources are affected by the water current. The node movement affects the sensor network design, the distance between the nodes (i.e., during route discovery), propagation, energy utilization among others. In an aquatic environment, electromagnetic wave propagation to sinks above water is challenging. Energy conservation and mobility in this type of network create challenges in the design of routing protocols. Routing protocols that are capable of minimizing the energy consumed by nodes, managing the random variation in the network topology and minimizing the delay in communication are therefore required in an aquatic environment.

In designing and evaluating mobility models for aquatic environments, our work took into consideration the conditions and characteristics of freshwater sources. Freshwater sources such as rivers are in constant motion. The velocity of the river depends on the slope of the land, the size and shape of the bed, and the quantity of water in the river [4, 32]. Sensor nodes may be deployed in one of the following environments; 1) slow but deep water bodies, 2) swift but deep water bodies, and 3) swift but shallow water bodies. Freshwater sources are characterized by three key factors: velocity, gradient, and discharge. Velocity is the distance that the water in a river travels in a given amount of time. The velocity relates to the energy levels of the water in the river. For example, objects (small or large) deposited in a swift or fast-moving river are carried downstream or along the river path quickly as compared to slow-moving rivers. The velocity of a river is affected by the gradient, discharge, and the shape of the river path that the water travels. The gradient is a measure of the steepness of a river, and a river's discharge is expressed as the quantity of water that moves through the different points along the river path at any given time. The discharge varies along the length of the river [32].

Mobile sensor nodes continue to move with water currents after the initial deployment. There are different mobility models used to simulate WSN projects. These include constant acceleration, constant position, constant velocity, Gauss Markov, hierarchical, random direction 2D, random waypoint, steady state random waypoint, random walk2D, and waypoint mobility models. Although these mobility


**Table 2.**

*Random mobility models compared.*

models are designed for terrestrial WSN projects, they may be modified for use in aquatic WSN projects. In considering the random walk mobility model, the mobile sensor node after the initial travel time stays at a point for a given period known as the pause time. In other models, the mobile sensor node chooses a new direction and a new speed throughout the travel time and travels towards the destination with this new direction and speed. A comparison of some random mobility models and their characteristics are presented in **Table 2**.

Random walk mobility model is designed in a way that the mobile sensor nodes move freely and randomly in the defined simulation field. In our application domain (i.e., river network monitoring), the mobile sensor node's movement is affected by the characteristics of the water environment. Notably, the movement is affected by the velocity of the water. Therefore, the sensor nodes are not expected to move in a


**251**

**Figure 7.**

*Energy conservation in wireless sensor networks.*

*Applications of Prediction Approaches in Wireless Sensor Networks*

• The node's directional movement is randomly generated

• The mobile sensor node moves independently of other nodes

that aids in the determination of a sensor nodes' mobility behavior.

Hence, there is a need to design models that have different mobility characteristics (i.e., such as destination, velocity, and direction). In an aquatic model defined in [33], the speed increases incrementally and the change in direction may not be smooth but fraught in some scenarios as shown in **Figure 7**. The Random Walk mobility model defined proposes node movements that depend on speed, discharge and gradient of the river. The speed and direction of the mobile sensor nodes are critical parameters

Energy conservation schemes in WSN include approaches that depend on characteristics of the data values produced in the network to mitigate energy consumption. Data-Driven techniques, shown in **Figure 8**, involve the use of data characteristics of a sampled data stream to mitigate energy consumption [27]. They include data reduction and energy efficient data acquisition. While data reduction schemes save energy by taking out the redundant data, energy efficient data acquisition reduces the energy spent by the sensing subsystem or sometimes through

Energy efficient data acquisition schemes work on the premise that the sensing subsystem, in some instances, consume significant amount of energy. These may include energy hungry transducers that require high power to sample data. Examples include multimedia sensors or biological sensor. Sensors that require active transducers like radar or laser rangers require active sensing and applications that require long acquisition time that may run into several hundred milliseconds or even seconds [33]. Reducing the energy consumption through frequent/continuous sensing translates into reducing the number of communication as well. The authors of [27, 34] argue that many energy efficient data acquisition schemes have been classified as methods to reduce the energy consumption of the radio. Therefore, WSNs assume the sensing subsystem has negligible energy consumption.

Data acquisition and reduction approaches have been used individually or simultaneously to mitigate energy consumption. The acquisition schemes measure data samples that are correlated spatially or temporally [35]. Temporally because subsequent data samples may not differ much from each other while data from

The inherent characteristics of the data that influence the data received at the sink may be classified as data reduction schemes. Data reduction approaches are generally categorized as In-networking processing, Data Compression and Data Prediction schemes. In-network processing applications are designed to be application specific. In-network data processing operations such as fusion and aggregation incorporate sensing and networking to reduce unnecessary data forwarding from

neighboring sensor nodes may not differ much in spatial-correlation.

*DOI: http://dx.doi.org/10.5772/intechopen.94500*

communication.

*Wireless Sensor Networks - Design, Deployment and Applications*

Node selects destination (uniformly distributed)

To specified destination in environment

Moves until it reaches the destination

Stays in a location for a period of time (Uniformly distributed)/Pause time

Stays in a location for a period of time (Uniformly distributed)/Pause time

Selects speed (uniformly distributed)

Selection of destination

Selection of destination

Selection of Speed

Moving to destination

Waiting time

Next destination

**Table 2.**

*Random mobility models compared.*

**Parameters Random waypoint Random walk Random** 

**Terrestrial environment Aquatic** 

Pre-defined destination (uniformly distributed)

Selects a direction (uniformly distributed)

Pre-defined speed (Uniformly distributed)

Node moves until reaching the boundary/a certain distance

Wait a certain amount of time

Node selects new direction after reaching boundary/ a certain distance

**Environment**

**walk**

**Modified random** 

Region of interest

Depends on water current (randomly chosen directions)

Depends on river current/speed

Nodes move to boundary and redirected by bouncing effects to change course

(x, y)

— Node floating to the

No waiting time since node movement is dependent on the flow motion

bank (Boundary) and bouncing effects determines (x, y) destination

**Direction**

MN travels to boundary of scenario (destination)

Selects uniform direction from boundary

Selects uniform speed

Moves until it reaches the boundary

Wait a certain amount of time

models are designed for terrestrial WSN projects, they may be modified for use in aquatic WSN projects. In considering the random walk mobility model, the mobile sensor node after the initial travel time stays at a point for a given period known as the pause time. In other models, the mobile sensor node chooses a new direction and a new speed throughout the travel time and travels towards the destination with this new direction and speed. A comparison of some random mobility models and

Random walk mobility model is designed in a way that the mobile sensor nodes

• The node's movement pattern is limited to an only small area in the simulation

move freely and randomly in the defined simulation field. In our application domain (i.e., river network monitoring), the mobile sensor node's movement is affected by the characteristics of the water environment. Notably, the movement is affected by the velocity of the water. Therefore, the sensor nodes are not expected to

• The time and distance that a mobile sensor node moves are short.

their characteristics are presented in **Table 2**.

**250**

move in a

environment


Hence, there is a need to design models that have different mobility characteristics (i.e., such as destination, velocity, and direction). In an aquatic model defined in [33], the speed increases incrementally and the change in direction may not be smooth but fraught in some scenarios as shown in **Figure 7**. The Random Walk mobility model defined proposes node movements that depend on speed, discharge and gradient of the river. The speed and direction of the mobile sensor nodes are critical parameters that aids in the determination of a sensor nodes' mobility behavior.

Energy conservation schemes in WSN include approaches that depend on characteristics of the data values produced in the network to mitigate energy consumption. Data-Driven techniques, shown in **Figure 8**, involve the use of data characteristics of a sampled data stream to mitigate energy consumption [27]. They include data reduction and energy efficient data acquisition. While data reduction schemes save energy by taking out the redundant data, energy efficient data acquisition reduces the energy spent by the sensing subsystem or sometimes through communication.

Energy efficient data acquisition schemes work on the premise that the sensing subsystem, in some instances, consume significant amount of energy. These may include energy hungry transducers that require high power to sample data. Examples include multimedia sensors or biological sensor. Sensors that require active transducers like radar or laser rangers require active sensing and applications that require long acquisition time that may run into several hundred milliseconds or even seconds [33]. Reducing the energy consumption through frequent/continuous sensing translates into reducing the number of communication as well. The authors of [27, 34] argue that many energy efficient data acquisition schemes have been classified as methods to reduce the energy consumption of the radio. Therefore, WSNs assume the sensing subsystem has negligible energy consumption.

Data acquisition and reduction approaches have been used individually or simultaneously to mitigate energy consumption. The acquisition schemes measure data samples that are correlated spatially or temporally [35]. Temporally because subsequent data samples may not differ much from each other while data from neighboring sensor nodes may not differ much in spatial-correlation.

The inherent characteristics of the data that influence the data received at the sink may be classified as data reduction schemes. Data reduction approaches are generally categorized as In-networking processing, Data Compression and Data Prediction schemes. In-network processing applications are designed to be application specific. In-network data processing operations such as fusion and aggregation incorporate sensing and networking to reduce unnecessary data forwarding from

**Figure 7.** *Energy conservation in wireless sensor networks.*

**Figure 8.** *Node mobility trajectory with random waypoint model.*

source nodes to the sink as they traverse the network. A survey of data aggregation methods was presented in 36], and present clustering-based aggregation protocols in WSN. The clustering protocols were classified as homogeneous, heterogeneous, and single and multi-hop.

Data compression is the reduction of the amount of data that traverses a network. It may be classified into lossless, loss and unrecoverable compression. In lossless compression, the data obtained after decompression is the same as data after the compression operation [37]. An example of lossless compression is the Huffman coding. In loss compression [38, 39], some details of the data are changed as a result of compression. Unrecoverable compression occurs when no decompression operation is applied such final data cannot be derived from the initial data [40, 41].

#### **6. Data prediction approaches**

Data prediction is a data reduction method in WSNs that reduces the number of data that is sent from sources to the sink. They are concerned with the building of models for forecasting sensed parameters of sensor nodes using historical data. The main aim of predictive wireless sensor networks is to exploit the temporal and spatial correlation characteristics of the observed environment. Prediction also capitalizes on the redundancies of environmental parameters to compute future parameter readings thereby reducing the frequency of required readings.

The forecasted values are adopted when they fall within some acceptable thresholds. Some predictive models are based on the sink and on the nodes themselves. Prediction is therefore performed on the sink node and since the same model is kept on the nodes by synchronization, it is assumed the same predicted values are generated. In some instances, prediction is done by the nodes and model is sent to the sink; hence the sink also computes the same predicted readings. When nodes take readings

**253**

**Figure 10.**

*Predictive approaches in wireless sensor networks.*

**Figure 9.**

*Data-driven approaches in WSN.*

*Applications of Prediction Approaches in Wireless Sensor Networks*

beyond some threshold, it is sent to the sink otherwise the sink uses the predicted values. When the model deteriorates such that the model does not reflect the current readings, fresh data is read from the environment and the model is updated. This requires frequent update and re-synchronization of the sensor nodes and the sink which introduces energy consumption overheads in the network [2]. To overcome the energy lost by synchronization, some predictive models are only kept on the nodes, which determine when readings from the sensors are transmitted to the sink. The prediction schemes surveyed in [22] on data reduction were not concerned with the medium access control challenges of radio energy consumption. In their classification, forecasting methods were not fully explored in WSN by the time of their publication as well as machine learning tools as argued in [42]. This is the basis for the new classification presented in **Figure 9**. Predictions studied in this chapter are limited to methods to predict data for the purposes for reducing data transmission (**Figure 10**).

Stochastic approaches characterize the sensed data as a random process from which a probabilistic model is applied to predict the data values. An example of a stochastic approach is proposed in [43] where KEN (a range of perception, understanding and knowledge based model) a robust approximation technique minimizes

*DOI: http://dx.doi.org/10.5772/intechopen.94500*

**6.1 Stochastic approaches**

#### *Applications of Prediction Approaches in Wireless Sensor Networks DOI: http://dx.doi.org/10.5772/intechopen.94500*

beyond some threshold, it is sent to the sink otherwise the sink uses the predicted values. When the model deteriorates such that the model does not reflect the current readings, fresh data is read from the environment and the model is updated. This requires frequent update and re-synchronization of the sensor nodes and the sink which introduces energy consumption overheads in the network [2]. To overcome the energy lost by synchronization, some predictive models are only kept on the nodes, which determine when readings from the sensors are transmitted to the sink.

The prediction schemes surveyed in [22] on data reduction were not concerned with the medium access control challenges of radio energy consumption. In their classification, forecasting methods were not fully explored in WSN by the time of their publication as well as machine learning tools as argued in [42]. This is the basis for the new classification presented in **Figure 9**. Predictions studied in this chapter are limited to methods to predict data for the purposes for reducing data transmission (**Figure 10**).

#### **6.1 Stochastic approaches**

*Wireless Sensor Networks - Design, Deployment and Applications*

source nodes to the sink as they traverse the network. A survey of data aggregation methods was presented in 36], and present clustering-based aggregation protocols in WSN. The clustering protocols were classified as homogeneous, heterogeneous,

Data compression is the reduction of the amount of data that traverses a network. It may be classified into lossless, loss and unrecoverable compression. In lossless compression, the data obtained after decompression is the same as data after the compression operation [37]. An example of lossless compression is the Huffman coding. In loss compression [38, 39], some details of the data are changed as a result of compression. Unrecoverable compression occurs when no decompression operation is applied such final data cannot be derived from the initial data [40, 41].

Data prediction is a data reduction method in WSNs that reduces the number of data that is sent from sources to the sink. They are concerned with the building of models for forecasting sensed parameters of sensor nodes using historical data. The main aim of predictive wireless sensor networks is to exploit the temporal and spatial correlation characteristics of the observed environment. Prediction also capitalizes on the redundancies of environmental parameters to compute future parameter readings thereby reducing the frequency of required readings.

The forecasted values are adopted when they fall within some acceptable thresholds. Some predictive models are based on the sink and on the nodes themselves. Prediction is therefore performed on the sink node and since the same model is kept on the nodes by synchronization, it is assumed the same predicted values are generated. In some instances, prediction is done by the nodes and model is sent to the sink; hence the sink also computes the same predicted readings. When nodes take readings

**252**

and single and multi-hop.

*Node mobility trajectory with random waypoint model.*

**Figure 8.**

**6. Data prediction approaches**

Stochastic approaches characterize the sensed data as a random process from which a probabilistic model is applied to predict the data values. An example of a stochastic approach is proposed in [43] where KEN (a range of perception, understanding and knowledge based model) a robust approximation technique minimizes

**Figure 10.** *Predictive approaches in wireless sensor networks.*

the communication from sensors to sink using replicated dynamic probabilistic models KEN exploits the spatial correlations across sensor nodes to boost coordination that minimizes communication cost. The Kalman filter, Dynamic Probabilistic Models and Monte Carlo are some algorithms that have been used in some stochastic models [38, 39]. Due to randomness and probabilities of data used in stochastic applications, they are usually energy consuming and care must be taken in applying them to WSNs. They are mostly used in heterogeneous networks were specialized nodes that have high energy levels run these models.

#### **6.2 Algorithmic approaches**

Algorithmic approaches are predictive methods that depend on algorithms that consider the heuristic or behavioral characteristics of the data collected from the environment. One of the first works based on the algorithmic approach is Prediction-based Monitoring (PREMON) [44]. PREMON is a prediction-based monitoring algorithm which is inspired by concepts from MPEG (also known as the Moving Pictures Experts Group [MPEG]), a standard for video compression. Algorithmic approaches are usually application specific. A spatiotemporal data map of data stream from sensors at the sink node is obtained by generating periodic snapshots at a given time granularity [45]. An Energy efficient data collection framework (EEDC) is proposed in [46] that exploits the spatiotemporal correlation of data to form clusters and a randomized scheduling algorithm to conserve energy. A lightweight greedy algorithm that minimizes the energy consumed through data transmission is modeled as an optimization problem called Piecewise Linear Approximation with minimum number of Line Segments (PLAMLiS). The algorithm is integrated into EEDC framework to further save energy. In [47] a data collection method is proposed that exploits the trade-off between QoS of applications and energy consumption by progressively exposing power saving states in a network while minimizing the energy consumption. Since algorithms are application specific, more work may be done in the future in real applications to standardize the approaches. In [48], the challenge of selecting the efficient prediction model for a particular time series is addressed by an adaptive lightweight and online model selection algorithm. The algorithm statistically selects an optimal model among a list of candidates. The Dual Prediction Scheme (DPS) has been used severally in WSN applications and yield high communication and energy savings.

### **6.3 Machine learning approaches**

Machine learning uses tools, techniques and algorithms with computational viability to improve on themselves by learning through their experience on data and is used for creating predictive models. They are able to carry out predictions while learning from streams of data patterns. Machine learning presents strategies for WSNs to better make informed decisions from the data in the network to improve network performance. These strategies learn the behavior of the networks and improve on its experiences of the environment without explicit programming. It is based on statistical models and probabilities to make predictions of future data occurrences.

In literature, machine learning approaches are categorized as supervised learning, unsupervised learning and reinforcement learning [Under supervised learning approaches such as K-nearest neighbor (k-NN), Decision Trees, Neural Networks, support vector machines and Bayesian statistics are the most commonly used Unsupervised learning techniques include K-Means clustering, and Principal Component Analysis (PCA). Reinforcement learning approaches have been used with the most common approach in WSN being Q-learning. Most of

**255**

**Mechanisms** Large scale network clustering

Location-aware activity recognition

Localization based on Neural Networks

Soft Localization

Cluster Head election

Underwater surveillance systems

Clustering using SOM and sink

Unsupervised learning

SOM

distance

Path Determination

Role-free clustering

Localization using SVM

**Table 3.**

*Feasibility of machine learning approaches in wireless sensor networks.*

Reinforcement

Reinforcement

Low

Learning

Q-learning

Low

No

Low

Low

No

Yes

Learning

Reinforcement

Learning

Supervised learning

SVM

Moderate

Supervised Learning

Decision Trees

Low Moderate Moderate

No

High

Moderate

Yes

Yes

Yes

Low

Low

Yes No

Supervised Learning

Neural Networks

High Moderate

Supervised Learning

Bayesian Networks

Moderate Moderate

Yes

Yes

High High

Low

Yes Yes

**Classification**

**Machine leaning algorithms**

**Complexity**

**Balancing energy consumption**

**Delay**

**Overhead**

**Topology**

*Applications of Prediction Approaches in Wireless Sensor Networks*

Yes Yes

*DOI: http://dx.doi.org/10.5772/intechopen.94500*


#### *Applications of Prediction Approaches in Wireless Sensor Networks DOI: http://dx.doi.org/10.5772/intechopen.94500*

*Wireless Sensor Networks - Design, Deployment and Applications*

nodes that have high energy levels run these models.

**6.2 Algorithmic approaches**

**6.3 Machine learning approaches**

the communication from sensors to sink using replicated dynamic probabilistic models KEN exploits the spatial correlations across sensor nodes to boost coordination that minimizes communication cost. The Kalman filter, Dynamic Probabilistic Models and Monte Carlo are some algorithms that have been used in some stochastic models [38, 39]. Due to randomness and probabilities of data used in stochastic applications, they are usually energy consuming and care must be taken in applying them to WSNs. They are mostly used in heterogeneous networks were specialized

Algorithmic approaches are predictive methods that depend on algorithms that consider the heuristic or behavioral characteristics of the data collected from the environment. One of the first works based on the algorithmic approach is Prediction-based Monitoring (PREMON) [44]. PREMON is a prediction-based monitoring algorithm which is inspired by concepts from MPEG (also known as the Moving Pictures Experts Group [MPEG]), a standard for video compression. Algorithmic approaches are usually application specific. A spatiotemporal data map of data stream from sensors at the sink node is obtained by generating periodic snapshots at a given time granularity [45]. An Energy efficient data collection framework (EEDC) is proposed in [46] that exploits the spatiotemporal correlation of data to form clusters and a randomized scheduling algorithm to conserve energy. A lightweight greedy algorithm that minimizes the energy consumed through data transmission is modeled as an optimization problem called Piecewise Linear Approximation with minimum number of Line Segments (PLAMLiS). The algorithm is integrated into EEDC framework to further save energy. In [47] a data collection method is proposed that exploits the trade-off between QoS of applications and energy consumption by progressively exposing power saving states in a network while minimizing the energy consumption. Since algorithms are application specific, more work may be done in the future in real applications to standardize the approaches. In [48], the challenge of selecting the efficient prediction model for a particular time series is addressed by an adaptive lightweight and online model selection algorithm. The algorithm statistically selects an optimal model among a list of candidates. The Dual Prediction Scheme (DPS) has been used severally in

WSN applications and yield high communication and energy savings.

Machine learning uses tools, techniques and algorithms with computational viability to improve on themselves by learning through their experience on data and is used for creating predictive models. They are able to carry out predictions while learning from streams of data patterns. Machine learning presents strategies for WSNs to better make informed decisions from the data in the network to improve network performance. These strategies learn the behavior of the networks and improve on its experiences of the environment without explicit programming. It is based on statisti-

cal models and probabilities to make predictions of future data occurrences. In literature, machine learning approaches are categorized as supervised learning, unsupervised learning and reinforcement learning [Under supervised learning approaches such as K-nearest neighbor (k-NN), Decision Trees, Neural Networks, support vector machines and Bayesian statistics are the most commonly used Unsupervised learning techniques include K-Means clustering, and Principal Component Analysis (PCA). Reinforcement learning approaches have been used with the most common approach in WSN being Q-learning. Most of

**254**

**Table 3.**

*Feasibility of machine learning approaches in wireless sensor networks.*

these approaches have been applied in sensor node clustering and data aggregation mechanisms as presented in the **Table 3**.

In [46], the authors present surveys of machine learning approaches that have been applied to WSN to address known challenges such as outlier detection in data, computational efficiency, and errors in collected data. They are used in tasks including classification and regression and applied in bioinformatics, computer science, spam detection, anomaly detection, and fraud detection. Machine learning approaches have provided a means for wireless sensor networks to dynamically adapt its behavior to the environment. They are used to solve WSNs challenges such as limited resources and the diversity of the learning themes and patterns developed from the vast data collected. In [46], the authors identified challenges with machine learning approaches in wireless sensor networks which include developing lightweight and distributed message passing techniques, online learning algorithms, hierarchical clustering pattern and to develop solutions that take into consideration the resource limitation of WSN.

#### **7. Sleep scheduling based on prediction**

Sleep scheduling algorithms reduce the energy consumption in wireless sensor nodes due to idle listening and overhearing. Traditionally, sleep scheduling algorithms require frequent sleep/wake-up transitions; however, they consume extra energy due to the frequent state transitions. Sleep scheduling algorithms determine some nodes to be awake in a given epoch, while the remaining nodes maximize their sleep periods to reduce the energy consumed in the network.

Several sleep scheduling approaches have been proposed in literature since nodes in sleep (low power) modes consume significantly lower energy than active nodes. Some scheduling algorithmic approaches in literature propose means of turning off redundant nodes in the network, leaving some active nodes to continue network operations. Initial works include [49] where longer sleep cycles were proposed for a network with redundant nodes. The authors explored the relationship between the level of redundancy and the low duty cycles. The approach synchronizes nodes for a robust performance for large networks as opposed to a random scheduling approach. But the cost of synchronizing sensor nodes is high when the control overhead increases with increasing size of the network.

Idle listening, which occurs when nodes stay awake for longer periods waiting for packets, is a major source of energy waste in WSN. Authors in [49, 50] suggests idle listening consumes just as much energy required for receiving packets. Hence nodes are turned off when the radio module is not in use. Several approaches have been introduced by authors to mitigate idle listening.

One such approach is to put nodes in low-power listening (LPL) low duty cycle (LDC) mode (when nodes are made to sleep for longer periods when there is no activity in the network) that makes nodes wake up periodically to check the channel for activity. Characteristics of LDC include networks in periodic and relatively low frequency of network activity. One challenge of such networks is synchronization of receiver and sender nodes. Since sections of the network may be turned-off to conserve energy, remote communication may result in packets not reaching their destination, or have increased end-to-end delays. An example is Berkeley-MAC (B-MAC) makes nodes independently sample the channel for activity, by sending a preamble longer than the listening time of the duty cycle of the receiver node. This ensures the receiver is awake to receive its intended message. But this increases energy waste when the load in the network increases and collisions occur as a result of the increased preambles.

**257**

*Applications of Prediction Approaches in Wireless Sensor Networks*

receive packets when transmitted from the sender nodes.

Low-power listening with Wake-Up After Transmission (LWT-MAC) was developed in [51] that alerts all nodes that overhead the last transmitted packets to wakeup to receive the new message. This approach does not depend on preambles but on the local synchronization of the nodes. This approach does not consider remote communication beyond the local nodes that overhear the latest transmission. Other protocols like X-MAC [52], S-MAC [42] and T-MAC [53] uses the Request-to-Send (RTS) and Clear-to-Send (CTS) to keep receiver nodes active when a message is intended for them. Other low-power listening protocols include Scheduled Channel Polling Mac (SCP-MAC) that increases the duty cycle after a successful message

TDMA protocols used in duty cycle algorithms reduce idle listening, over hearing and also avoid collisions. But they also increase the number of dropped packets in transmissions since nodes may only receive packets when they are active. Synchronization of receiver nodes and sender nodes ensures nodes are active to

The medium access therefore is one key challenge in the era of Internet of Things (IoTs) due to the huge number of device connectivity with different traffic profiles. It is therefore, of necessity, an urgent requirement of these sensor nodes to reduce the number of transmissions from devices using data reduction approaches. However, prediction models, which forecast future data values may be a promising tool to reduce transmissions while maintaining optimum levels of trust and reliabil-

In the past, researchers avoided the implementation of complex algorithms in WSN due to their limited computational capabilities. However, with recent advances in hardware, data mining techniques are increasing used in WSN to discover patterns in sensor data to improve on its successful delivery [54]. Even though not all data mining approaches include predictions, machine learning tools that implement prediction in WSN is currently an open research area. Authors in [54] discussed the implementation of machine learning in the routing and medium

In the implementation of prediction models, time series models have been used together with machine learning to forecast data values. This is because, machine learning-based approaches uncover and learn patterns in data to evolve their predictions in response to changes in the deployed environment. On the other hand, time series methods depend on the statistical and probabilistic tendencies of the data to make predictions. The survey by [55] introduced current prediction techniques for

Time series models sequentially model observed data over successive time. The successive data is analyzed to extract meaningful statistics and probabilities that may influence calculated forecasts. Time series data may be seasonal, trend, cyclical or irregular. Trend analysis tends to increase, decrease or remain stagnant over a long period. Seasonal analysis shows fluctuations that change during repeatable periods in the period under observation. Under cyclical, the patterns observed describe medium-term changes that may repeat in cycles, where cycles may extend over long periods. Irregular patterns fluctuate unpredictably and hence are difficult

Examples of time series models implemented in WSN include the naive approaches, autoregressive and moving average approaches (AR and MA), Auto-Regressive Integrated Moving Average (ARIMA), Exponential Smoothing, Gray Series, and Least Mean Square (LMS). In time series analysis, Autoregressive Integrated Moving Average (ARIMA) generalizes the Autoregressive moving average (ARMA) with a differencing. However, ARMA is a combination of the Autoregressive (AR) and Moving Average (MA) models into a single equation.

*DOI: http://dx.doi.org/10.5772/intechopen.94500*

delivery to cut down on end-to-end delays.

ity of the data received [53].

reducing data transmission in WSN.

to be predicted using time series methods.

access layers.

#### *Applications of Prediction Approaches in Wireless Sensor Networks DOI: http://dx.doi.org/10.5772/intechopen.94500*

*Wireless Sensor Networks - Design, Deployment and Applications*

mechanisms as presented in the **Table 3**.

the resource limitation of WSN.

**7. Sleep scheduling based on prediction**

sleep periods to reduce the energy consumed in the network.

overhead increases with increasing size of the network.

been introduced by authors to mitigate idle listening.

these approaches have been applied in sensor node clustering and data aggregation

In [46], the authors present surveys of machine learning approaches that have been applied to WSN to address known challenges such as outlier detection in data, computational efficiency, and errors in collected data. They are used in tasks including classification and regression and applied in bioinformatics, computer science, spam detection, anomaly detection, and fraud detection. Machine learning approaches have provided a means for wireless sensor networks to dynamically adapt its behavior to the environment. They are used to solve WSNs challenges such as limited resources and the diversity of the learning themes and patterns developed from the vast data collected. In [46], the authors identified challenges with machine learning approaches in wireless sensor networks which include developing lightweight and distributed message passing techniques, online learning algorithms, hierarchical clustering pattern and to develop solutions that take into consideration

Sleep scheduling algorithms reduce the energy consumption in wireless sensor nodes due to idle listening and overhearing. Traditionally, sleep scheduling algorithms require frequent sleep/wake-up transitions; however, they consume extra energy due to the frequent state transitions. Sleep scheduling algorithms determine some nodes to be awake in a given epoch, while the remaining nodes maximize their

Several sleep scheduling approaches have been proposed in literature since nodes in sleep (low power) modes consume significantly lower energy than active nodes. Some scheduling algorithmic approaches in literature propose means of turning off redundant nodes in the network, leaving some active nodes to continue network operations. Initial works include [49] where longer sleep cycles were proposed for a network with redundant nodes. The authors explored the relationship between the level of redundancy and the low duty cycles. The approach synchronizes nodes for a robust performance for large networks as opposed to a random scheduling approach. But the cost of synchronizing sensor nodes is high when the control

Idle listening, which occurs when nodes stay awake for longer periods waiting for packets, is a major source of energy waste in WSN. Authors in [49, 50] suggests idle listening consumes just as much energy required for receiving packets. Hence nodes are turned off when the radio module is not in use. Several approaches have

One such approach is to put nodes in low-power listening (LPL) low duty cycle (LDC) mode (when nodes are made to sleep for longer periods when there is no activity in the network) that makes nodes wake up periodically to check the channel for activity. Characteristics of LDC include networks in periodic and relatively low frequency of network activity. One challenge of such networks is synchronization of receiver and sender nodes. Since sections of the network may be turned-off to conserve energy, remote communication may result in packets not reaching their destination, or have increased end-to-end delays. An example is Berkeley-MAC (B-MAC) makes nodes independently sample the channel for activity, by sending a preamble longer than the listening time of the duty cycle of the receiver node. This ensures the receiver is awake to receive its intended message. But this increases energy waste when the load in the network increases and collisions occur as a result

**256**

of the increased preambles.

Low-power listening with Wake-Up After Transmission (LWT-MAC) was developed in [51] that alerts all nodes that overhead the last transmitted packets to wakeup to receive the new message. This approach does not depend on preambles but on the local synchronization of the nodes. This approach does not consider remote communication beyond the local nodes that overhear the latest transmission. Other protocols like X-MAC [52], S-MAC [42] and T-MAC [53] uses the Request-to-Send (RTS) and Clear-to-Send (CTS) to keep receiver nodes active when a message is intended for them. Other low-power listening protocols include Scheduled Channel Polling Mac (SCP-MAC) that increases the duty cycle after a successful message delivery to cut down on end-to-end delays.

TDMA protocols used in duty cycle algorithms reduce idle listening, over hearing and also avoid collisions. But they also increase the number of dropped packets in transmissions since nodes may only receive packets when they are active. Synchronization of receiver nodes and sender nodes ensures nodes are active to receive packets when transmitted from the sender nodes.

The medium access therefore is one key challenge in the era of Internet of Things (IoTs) due to the huge number of device connectivity with different traffic profiles. It is therefore, of necessity, an urgent requirement of these sensor nodes to reduce the number of transmissions from devices using data reduction approaches. However, prediction models, which forecast future data values may be a promising tool to reduce transmissions while maintaining optimum levels of trust and reliability of the data received [53].

In the past, researchers avoided the implementation of complex algorithms in WSN due to their limited computational capabilities. However, with recent advances in hardware, data mining techniques are increasing used in WSN to discover patterns in sensor data to improve on its successful delivery [54]. Even though not all data mining approaches include predictions, machine learning tools that implement prediction in WSN is currently an open research area. Authors in [54] discussed the implementation of machine learning in the routing and medium access layers.

In the implementation of prediction models, time series models have been used together with machine learning to forecast data values. This is because, machine learning-based approaches uncover and learn patterns in data to evolve their predictions in response to changes in the deployed environment. On the other hand, time series methods depend on the statistical and probabilistic tendencies of the data to make predictions. The survey by [55] introduced current prediction techniques for reducing data transmission in WSN.

Time series models sequentially model observed data over successive time. The successive data is analyzed to extract meaningful statistics and probabilities that may influence calculated forecasts. Time series data may be seasonal, trend, cyclical or irregular. Trend analysis tends to increase, decrease or remain stagnant over a long period. Seasonal analysis shows fluctuations that change during repeatable periods in the period under observation. Under cyclical, the patterns observed describe medium-term changes that may repeat in cycles, where cycles may extend over long periods. Irregular patterns fluctuate unpredictably and hence are difficult to be predicted using time series methods.

Examples of time series models implemented in WSN include the naive approaches, autoregressive and moving average approaches (AR and MA), Auto-Regressive Integrated Moving Average (ARIMA), Exponential Smoothing, Gray Series, and Least Mean Square (LMS). In time series analysis, Autoregressive Integrated Moving Average (ARIMA) generalizes the Autoregressive moving average (ARMA) with a differencing. However, ARMA is a combination of the Autoregressive (AR) and Moving Average (MA) models into a single equation.


*Time series classification.*

**259**

**8. Conclusion**

*Applications of Prediction Approaches in Wireless Sensor Networks*

These models are best for stochastic data sets that may be linear or non-linear. ARIMA models have several applications in WSN due to their simplicity of use, and

Exponential smoothing prediction is a time series forecasting method that can be used to forecast univariate data to support data with systemic or seasonal trends [56]. It is a model where predictions are weighted sums of past observations, such that weights exponentially decrease as they get older. Weights of predictions decrease at a geometric ratio. They may be considered peers or alternatives to other Box-Jenkins models. Three exponential smoothing models are generally used, that is the single, double and triple exponential smoothing models. The single exponential smoothing methods add no systemic structure to the univariate data without trend or seasonality. A smoothing factor, alpha is set to values between 0 and 1, with larger values indicating model is dependent more on recent observations and the vice versa. The single smoothing method is used when the observed data is either

The double smoothing method adds trends and seasonality to the univariate data. A beta value is added to the alpha value to control the decay of the influence of change in the trend of data. The change may be additive or multiplicative. When the change is additive, the Holt's linear trend model. In longer forecasts, the double smoothing method may reduce the size of the trend to straighten it, prevent unrealistic trends. In the triple smoothing, a gamma value is added to the alpha and beta to

Gray series are time series prediction model with superiority to other statistical methods of prediction [57] due to its accuracy in predicting small data sets with lower errors. Gray models are preferable for data where statistical methods may not be appropriate or if the data does not satisfy conventional distributions [57]. In WSN, gray series are preferred due to their minimum complexity and low computational capacity. **Table 4** present some selected classification of time series models

The chapter focuses on the various deployment environments and the challenges that the implementation of WSNs brings about when adopted for use. The chapter also discussed WSN topologies and explored ways in which network topology impacts energy consumption and communication issues. The chapter highlights the different techniques mainly used when implementing WSNs to maximize lifetime, coverage, and connectivity while ensuring data fidelity. The chapter also describes the energy conservation techniques, data prediction approaches, and mobility models. Various data prediction models such as the time series models are discussed in detail. A classification of some time series models based on their popularity and recent use highlights their applications. Advantages and disadvantages of the selected time series models and reasons of their implementations are discussed.

*DOI: http://dx.doi.org/10.5772/intechopen.94500*

near accurate prediction for linear data.

stationary or changes slowly with time.

influence the seasonality of the observed data.

used in WSN based on their popularity and frequency of use.

#### *Applications of Prediction Approaches in Wireless Sensor Networks DOI: http://dx.doi.org/10.5772/intechopen.94500*

These models are best for stochastic data sets that may be linear or non-linear. ARIMA models have several applications in WSN due to their simplicity of use, and near accurate prediction for linear data.

Exponential smoothing prediction is a time series forecasting method that can be used to forecast univariate data to support data with systemic or seasonal trends [56]. It is a model where predictions are weighted sums of past observations, such that weights exponentially decrease as they get older. Weights of predictions decrease at a geometric ratio. They may be considered peers or alternatives to other Box-Jenkins models. Three exponential smoothing models are generally used, that is the single, double and triple exponential smoothing models. The single exponential smoothing methods add no systemic structure to the univariate data without trend or seasonality. A smoothing factor, alpha is set to values between 0 and 1, with larger values indicating model is dependent more on recent observations and the vice versa. The single smoothing method is used when the observed data is either stationary or changes slowly with time.

The double smoothing method adds trends and seasonality to the univariate data. A beta value is added to the alpha value to control the decay of the influence of change in the trend of data. The change may be additive or multiplicative. When the change is additive, the Holt's linear trend model. In longer forecasts, the double smoothing method may reduce the size of the trend to straighten it, prevent unrealistic trends. In the triple smoothing, a gamma value is added to the alpha and beta to influence the seasonality of the observed data.

Gray series are time series prediction model with superiority to other statistical methods of prediction [57] due to its accuracy in predicting small data sets with lower errors. Gray models are preferable for data where statistical methods may not be appropriate or if the data does not satisfy conventional distributions [57]. In WSN, gray series are preferred due to their minimum complexity and low computational capacity. **Table 4** present some selected classification of time series models used in WSN based on their popularity and frequency of use.
