- Email: [email protected]
An energy prediction algorithm for wind-powered wireless sensor networks with energy harvesting
Accepted Manuscript An Energy Prediction Algorithm for Wind-Powered Wireless Sensor Networks with Energy Harvesting
Selahattin Kosunalp PII:
S0360-5442(17)30965-9
DOI:
10.1016/j.energy.2017.05.175
Reference:
EGY 10983
To appear in:
Energy
Received Date:
30 December 2016
Revised Date:
24 May 2017
Accepted Date:
28 May 2017
Please cite this article as: Selahattin Kosunalp, An Energy Prediction Algorithm for Wind-Powered Wireless Sensor Networks with Energy Harvesting, Energy (2017), doi: 10.1016/j.energy. 2017.05.175
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
An Energy Prediction Algorithm for Wind-Powered Wireless Sensor Networks with Energy Harvesting Selahattin Kosunalp 1Department
of Industrial Engineering, Faculty of Engineering, University of Bayburt, 69000 Bayburt, Turkey 2Central Research Laboratory, University of Bayburt, 69000 Bayburt,
Email: [email protected]
Abstract— Energy harvesting (EH) from environmental energy sources has the potential to ensure unlimited, uncontrollable and unreliable energy for wireless sensor networks (WSNs), bringing a need to predict future energy availability for the effective utilization of the harvested energy. The majority of previous prediction approaches have exploited the diurnal cycle dividing the whole day into equal-length time slots in which predictions were carried out in each slot independently. This is not, however, efficient for wind energy as it exhibits non-controllable behaviour in that the amount of energy to be harvested varies over time. This paper proposes a novel approach to predict the wind energy for EH-WSNs depending on the energy generation profile of latest condition. The distinctive feature of the proposed approach is to consider the recent conditions in current-day, instead of past-day’s energy generation profiles. The performance of the proposed algorithm is evaluated using real measurements in comparison with state-of-art approaches. Results show that the proposed strategy significantly outperforms the two popular energy predictors, EWMA and ProEnergy. Index Terms— Energy harvesting, wind power, wireless sensor networks, energy prediction.
1. Introduction The main body of wireless sensor networks (WSNs) includes a collection of resourceconstrained sensor nodes performing a common task [1]. A typical sensor node senses an environmental parameter, such as temperature or movement, which is then processed to be transmitted to a central point, called sink or base-station. Sensor nodes are solely powered by limited small batteries as depicted in Fig. 1 with all components. Traditionally, sensor nodes are randomly deployed in harsh and remote areas where a physical access to the location of the sensor nodes is often impractical resulting in recharging or replacement of the battery a difficult task. The energy constraints of sensor nodes may limit the functions of the WSNs. A
ACCEPTED MANUSCRIPT successful deployment of long-lasting WSNs relies on the effective utilization of the current energy level. The emphasis has therefore been placed on maximizing the operation lifetime of WSNs. Wireless communication consumes the majority of energy source when compared to energy consumptions in sensing and processing parts. Therefore, an intelligent coordination of channel capacity among sensor nodes is required to effectively manage the energy [2]. It is well-understood that energy is wasted through idle-listening, overhearing, extra control packet exchange and collisions [3]. Medium access control (MAC) protocols are necessarily designed to minimize aforementioned energy wastage reasons through assigning intelligent capacity of channel and coordinating communication in collision-free manner [4, 5]. A huge number of MAC protocols have already been proposed for WSNs with the core responsibility of energy-efficiency [6-8]. However, the energy will eventually deplete resulting in nodes losing their operation. In order to handle the bottleneck of energy in all energy-limited systems, renewable energy sources in the environment can be exploited as an unlimited energy source [9, 10]. General principle of this idea is to convert an existing ambient energy into electricity supplying the energy burden of the systems. Therefore, EH process from the surrounding environment would potentially provide an infinite energy source through continuous energy harvesting. Sensor nodes with the capability of EH exploit the ambient energy by employing an EH unit [11] as presented in Fig. 2. This unit includes an energy harvester, such as a solar panel or a wind turbine, and an energy storage component to accumulate the harvested energy. As energy is infinite, sensor nodes can operate perpetually, subject to hardware failure. Therefore, traditional MAC protocols may not be applicable to WSNs with EH (EH-WSN). This paradigm opens a new forum to the researchers. The fundamental design criterion has shifted from energy-efficiency to maximizing utilization of the energy harvested [12]. A number of MAC protocols explicitly designed for EH-WSNs have been surveyed in [13]. The proposed protocols considers the current rate of energy harvesting and the amount of energy stored. Another way of powering sensor nodes is RF energy transfer which relies on broadcasting energy packets by an energy emitting node [14]. In most cases, the energy packets are periodically sent to avoid instant energy shortages. However, periodic transfer of same level of energy would not be efficient as energy requirement of each node may exhibits a varying-characteristic. One of the major issues in RF energy transfer is the use of same frequency for both data and energy transfer. This brings necessity to carefully arrange the
ACCEPTED MANUSCRIPT sessions of data and energy. Also, the number of energy emitting nodes should be appropriately adjusted to particularly cover large areas. In order to solve these problems, new MAC protocols have been gaining much attention [15]. Solar and wind energy are the most effective energy source because of their high power intensity. The amount of energy to be harvested is time-varying depending on the environmental conditions. An accurate prediction of future energy availability is therefore of paramount importance. Most of the proposed energy prediction algorithms take solar energy into consideration. This is because the prediction of solar energy is relatively straightforward when compared to other forms of ambient energy. Existing approaches exploits the diurnal cycle of sun and assumes that the amount of energy harvested in a particular time slot of a day is similar to the energy harvested in the same time slot of the following days [16]. Basically, these approaches divides a day into equal-length time slots and predicts the energy independently in each slot. Due to the significant prediction errors when the environmental or weather conditions change frequently, a number of prediction approaches considers the current conditions in order to minimize the prediction errors [16-21] and these schemes are introduced in detail in the following section. Sensor nodes are often required to perpetually operate and an energy shortage would interrupt the operation of a typical WSN. The level of solar energy to be harvested in a particular time depends highly on intensity of sun. It is obvious that there will be no energy to harvest during night-time. The time of darkness takes long hours and sensor nodes must use their accumulated energy. A sensor node would potentially deplete its energy in night-time resulting in losing its operation. Therefore, the node becomes useless until energy harvesting begins, so that some important information may be lost. The main advantage of wind energy is its existence for a whole day. This will give sensor nodes more chances to stay alive in nigh-time. Nevertheless, prediction of wind energy is an important issue to establish. The trend of wind energy is more likely to change from day to day. Unlike solar energy, the predictions do not necessarily consider the diurnal cycle for an accurate estimation of wind energy. It is believed that recent information about the wind energy would carry more important information about the near-future energy prediction. This paper presents a new algorithm that predicts the future energy availability for wind energy. The performance of the proposed algorithm is compared with state-of-art approaches. In order to reflect the actual performance of the approaches, we obtained real measurements from two different sources which cover the period of 1-year in 2015. Section II presents the
ACCEPTED MANUSCRIPT description of state-of-art approaches covered in this paper and the fundamentals of the proposed algorithm. Performance outputs of the proposed approach in comparison to other approaches are presented in Section III. Section IV concludes the paper.
2. Existing Energy Prediction Approaches This section presents the details of the energy prediction algorithms which are compared with the proposed algorithm. The first scheme, called exponentially-weighted moving average (EWMA), can be accepted as a baseline scheme as many prediction algorithms have used the theme of EWMA [16]. It introduces the concept of equal-length time slots in which the energy prediction is explicitly performed in each slot. Weather-conditioned moving average (WCMA) improves the performance of EWMA considering not only the past day observation, but also the conditions in the present day [18]. PROfile Energy (Pro-Energy) targets at forecasting both solar and wind power [19]. Similar to EWMA, a day is considered as a collection of time slots. Then, the proposed algorithm is described in details.
2.1 Exponentially-weighted moving average (EWMA) The basic idea of EWMA, as discussed previously, is to split the day into identical time slots in terms of slot duration [16]. It assumes that the energy generation pattern in consecutive days exhibits a similar behaviour. Therefore, EWMA uses historical information of the energy generation pattern as presented in Equation (1). For this purpose, the last amount of harvested energy (H) and estimated energy (E) by EWMA are summed with a weighting factor, 0<α<1, arranging the importance of the H and E. A high value of α corresponds to less importance of the last-harvested energy and vice versa. In this equation, d represents the present day and n is the slot identifier. E(d, n) = α E(d-1, n) + (1 – α) H(d-1, n)
(1)
A distinctive feature of EWMA is its adaptation to seasonal variations through a weighted average of the harvestable energy in each slot in association with a set of energy available in past days. In order to mitigate such seasonal change, the value of α should be accurately arranged. In the original paper of EWMA, α value of 0.5 has been experimentally validated as an optimum value minimizing the prediction errors. A typical example of EWMA with 24 slots in adapting to a seasonal change using three different α values is presented in Fig. 3. This example shows the behaviour of EWMA process with various α values in a particular slot when the amount of harvested energy suddenly becomes four times greater than the expected energy. It is assumed that the harvested energy continues to be the same value. Due
ACCEPTED MANUSCRIPT to the higher real energy measurement, smaller α values cause the prediction to converge to the real energy measurement more aggressively. Therefore, the choice of α value becomes a crucial issue to establish which impacts on the time efficiency of the adaptation to seasonal changes.
2.2 Weather-conditioned moving average (WCMA) WCMA has it basics on EWMA with the introduction of current weather conditions in order to account for the influence of weather conditions of the current day [17]. This improvement essentially reduces the deficiencies of the EWMA in particularly frequently changing weather conditions. WCMA maintains a collection of past days energy values in a matrix E(i, j), where j indicates a particular slot on the ith day. When estimating an energy value in a slot, WCMA collects the amount of energy harvested in the previous slot, the average energy value of the relevant slot for a pre-defined number of previous days and a weighting factor representing the current solar conditions: (2)
E(d, n) = α H + (1 – α) MD(d, n)GAP
Here, α is the weighting factor as defined in EWMA. H carries the last amount of energy harvested and MD is the average of the sample for D previous days. The main contribution in this scheme is the GAP value to measure the current solar conditions. A mean value of a sample (n) on the dth day is calculated for D past days as: d−1
M(d, n) =
�
E(𝑖, 𝑛)
𝑖=𝑑−D
(3)
D General principle of GAP value is to observe the trend of the previous slots with respect to the previous days. To do this, the behaviour of solar energy outputs in K number of previous slots is considered. In order to calculate the GAP value, a vector with K elements, V = [V1, V2,…, VK], is initially defined to contain a numerical representation of energy generation profile in each K past slots. Each value of V is actually the ratio of the energy harvested to the average value: Vk = E(d, n-K+k) M(d, n-K+k)
(4)
ACCEPTED MANUSCRIPT It is obvious that the information in the recent slots would carry more important news about the current status of weather. WCMA assigns more importance to the recent slots through an another vector, P = [P1, P2, …, PK], where each element is defined according to the distance to the actual slot: Pk = k K
(5)
Once the values of the V vector are obtained, the GAP value is calculated as: GAP = V.P
(6)
∑P
2.3 Profile energy prediction model (Pro-Energy) PROfile energy (Pro-Energy) also discretizes each day into time slots and energy prediction for each time slot is derived at the beginning of the associated time slot [18]. Similar to WCMA, Pro-Energy utilizes the amount of energy harvested in the last slot. The main concept is to maintain the energy generation profiles of D past days in a matrix, E(i, j). This matrix is then used to find out the most similar day to the current day in terms of energy generation similarity. Therefore, Pro-Energy combines the last energy harvested and the most similar energy profile from the recent past as presented in equation 7. Ê(d, n) = α H + (1 – α) EMS
(7)
Where H is the last energy harvested and Ems represents the amount of energy harvested in slot n of the most similar day. Again, α is the weighting factor as in EWMA and WCMA. A critical issue is the determination of the most similar day. The mean absolute error (MAE) of the D previous days is compared with the current day to find the most similar day, in which the day with the minimum MAE is pointed as the most similar day. The stored profile is updated with the end of a day. One typical way is to directly add the ending-day to the E matrix with the deletion of the farthest day. This is however not efficient in some cases, such as two similar days would be unnecessarily stored. Therefore, Pro-Energy performs two criterions for the decision of updating the E matrix with the current day: (1) a stored profile can stay in the matrix for a period of A days. The current day can be replaced with a day that has stayed more than A days and (2) if two or more similar days are detected, the most similar day to the current day is selected to be replaced.
ACCEPTED MANUSCRIPT For the sake of further improving the accuracy of the predictions, Pro-Energy proposes to collect the energy values from multiple profiles reducing the potential poor performance caused by a single profile. Each previous day contributes to predict the energy value with a weight factor depending on the MAE. This changes the term of the most similar day (Ems) with the weighted profile (WP). This is then computed to replace with Ems in equation 7. P
WP
w j 0
j
Esj
P 1
(8)
Where Esj represent each previous day based on their MAE and wj is computed as: wj 1
MAE Esj , C P
MAE E j 1
sj
,C
(9)
As a result, the final energy prediction equation with the multiple profile is given as: Ê(d, n) = α H + (1 – α) WP
(10)
2.4 Proposed Algorithm It may be noted that a constant value of α is the main drawback of the EWMA and WCMA resulting in a high vulnerability to frequently changing conditions. EWMA can potentially produce highly inaccurate predictions following a series of sunny and cloudy days. WCMA takes the current conditions into consideration in order to handle the main drawback of EWMA and provides an average prediction error ratio of about 20% better than EWMA. However, WCMA considers the value of energy obtained in the previous slot which leads to high prediction errors particularly at sunrise and sunset (or when wind starts and stops). ProEnergy attempts to further improve the deficiencies of EWMA and WCMA, exploiting the information obtained in the previous slot as in WCMA and finding the most similar previous day in terms of energy harvesting profile. It has been experimentally shown that Pro-Energy outperforms EWMA and WCMA significantly. Similar to WCMA, using the energy value in the previous slot would cause high prediction errors. The process of exploring the most similar previous day for a particular slot is performed through average prediction error ratios in the previous slots of past days, in which the day with the lowest error ratio is selected as the most similar day. Therefore, Pro-Energy has no mechanism to check the current conditions and is prone to finding the wrong previous energy profile. It is believed that
ACCEPTED MANUSCRIPT predictions on a daily basis would limit the prediction accuracy especially in frequently changing conditions. In this section, we introduce a novel wind energy prediction strategy for EH-WSNs, called wind energy predictor (WEP) that focuses on historical information of the latest energy generation, instead of past day’s energy generation profiles. Unlike solar energy, wind energy can be generated at any time as long as there is wind. Estimating the wind energy is usually a more difficult task, because average wind speed at a location depends on many environmental factors (e.g. trees and buildings). It is therefore possible to observe varying average wind speeds in near parts of a location. Diurnal cycle for wind energy may not be feasible in harsh environments. Instead of splitting the day into time slots, WEP generally divides the time into equal-length slots. Then, it estimates the energy of current time slot by the means of a certain number of previous time slots as depicted in Fig. 4. The main feature of WEP is to consider a certain number (N) of energy generation of previous slots. A weighted average of the N previous slots is calculated in association with the increasing index (i). An energy prediction in a single time slot is therefore carried out as in Equation (11) below. N
∑ H(i)*i
(11)
i=1
∑i H represents the actual harvested energy which is multiplied with increasing index (i) in order to give more importance to the closer time slots as most recent slots would carry the most recent information as discussed in WCMA. In order to further reduce the prediction error, we also consider the amount of energy prediction error in each of previous slots. Our motivation is to take the prediction accuracy in the previous slots into consideration for forecasting energy in a slot. To do this, an average amount of prediction error in N previous slot is added to the Eq. 11 as formed below: N
N
∑ H(i)*i
∑ H(i) – P(i)
i=1
(12)
+ i=1
∑i
N
Here, P is the predicted value in slot i. Therefore, energy prediction in a particular slot is related to the amounts of energy harvested and prediction error in previous slots.
ACCEPTED MANUSCRIPT 3. Performance Evaluation This section presents the performance of WEP in comparison to that of two energy predictors, EWMA and Pro-Energy. In all experiments of EWMA and Pro-Energy, we set the number of time slots to 24, so that a complete day is represented by 24 time slots, each of which corresponds to 1 hour duration. We did not introduce WCMA in the performance comparisons because Pro-Energy has better performance than WCMA. Another important parameter to be established is the weighting factor (α). It is set to 0.5 as this value is suggested to provide the minimum error ratio in the original papers of two approaches. ProEnergy observes a certain number of previous days to find the most similar energy generation pattern and a certain number of previous slots to decide the most similar days. As suggested in the original paper, it is set to 10 days and 7 slots, respectively. For WEP, the N is set 5 slots as it provides the best result. In order to reflect the actual performance of the approaches, we obtained real measurements from RITM project which covers the period of 1-year in 2015. RITM project is an official project established by the General directorate of Renewable Energy, the Republic of Turkey Ministry of Energy and Natural Resources [22]. It aims to develop a wind power monitoring system across Turkey. The second real-life traces of the harvested energy was obtained from the US National Renewable Energy Laboratory (NREL) [23] which is a publicly-available dataset. The main motivation behind choosing two datasets is to test the performance of the schemes in a non-frequently and frequently changing conditions, RITM and NREL respectively. Table I and II present the average prediction error ratios for three schemes over 1-year period in RITM and NREL datasets. EWMA has the worst error ratio because it is prone to significant errors when wind conditions change frequently. EWMA has no mechanism to check the conditions in the current day, so that it produces high errors in temporary changes. Pro-Energy has a better performance than EWMA as it estimates the energy based on the latest energy generation and the days with the most similar energy generation profile. Finding the most similar profile is performed to compare the total prediction errors among the previous days. Therefore, although the most similar day have the lowest total prediction error, a number of slots may experience significantly wrong energy profile. This can cause high prediction errors. In RITM, WEP has a two-time better performance than Pro-Energy as it only detects the latest conditions. It is concluded that WEP provides the best performance
ACCEPTED MANUSCRIPT with an average error ratio of 0.086 (~%8.6). WEP is able to detect temporal changes quickly within N number of slots. WEP also allows the closer slots to have more impact on the estimation of current time slot. In NREL, however, three schemes have worse performance as the wind conditions change very often. Nevertheless, WEP produces the most successful predictions by only considering the current conditions. Pro-Energy struggles to find the most similar past day as the energy generation pattern in consecutive days may vary significantly. Therefore, the process of performing the most similar day which uses the information of past days prediction errors as discussed above may be inefficient in highly-changing conditions. In both cases, WEP improves the prediction accuracy at a significant rate. The average prediction error ratios for a specific number of points (all points in our scenario) presented in Table I and Table II are calculated as: N
Error =
𝟏 𝐍
∑ │𝟏 ‒ 𝐇𝐏│
(13)
i=1
where H represents the amount of harvested energy in the associated slot and P is the predicted energy value from the approaches. In order to depict what the three schemes predict, we present the energy prediction accuracy for a random period in Fig. 5. Here, the amount of actual harvested energy and estimations of the three schemes are shown. As discussed above, Pro-Energy produces highly inaccurate predictions in some slots due to the wrong choice of the most similar previous day. EWMA has an exponential trend as expected producing significantly inaccurate predictions. WEP usually predicts correctly with an overall error ratio.
4. Conclusions This paper has presented a novel wind energy prediction approach (WEP) for energyharvesting wireless sensor networks. Contrary to existing studies, WEP only considers the current profile of wind energy generation while others use historical information of energy generation from past days. The performance of WEP has been tested using two real-life traces of the harvested energy. Performance results proved that WEP outperforms EWMA and Pro-Energy with a significant improvement. The results show that WEP reduces the prediction error ratio to 8.6 % from 55.9 % (EWMA) and 17.9% (Pro-Energy) in the first
ACCEPTED MANUSCRIPT dataset. For the second dataset which has a more highly variable pattern than the first, all schemes have a higher prediction error ratio and WEP, Pro-Energy and EWMA have error ratios of 43.6%, 75.5% and 125% respectively. Therefore, WEP exhibits a robust mechanism for providing more accurate predictions in frequently-changing conditions. ACKNOWLEDGEMENTS We would like to thank in advance the General directorate of Renewable Energy, the Republic of Turkey Ministry of Energy and Natural Resources, and US National Renewable Energy Laboratory (NREL) for providing real-measurement outputs. This work has been supported by Bayburt University Scientific Research and Development Support Program (BAP) in Turkey. The project number is 2016/01-11. References [1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam and E. Cayirci, “A survey on sensor networks,” Communications Magazine, 40, pp. 102-114, 2002. [2] C. Gherbi, Z. Aliouat and M. Benmohammed, “An adaptive clustering approach to dynamic load balancing and energy efficiency in wireless sensor networks,” Energy, 114, pp. 647-662, 2016. [3] S. Kosunalp, P.D. Mitchell, D. Grace and T. Clarke, “Experimental study of capture effect for medium access control with ALOHA,” ETRI Journal, 37(2), pp. 359-368, 2015. [4] S. Kosunalp, Y. Chu, P.D. Mitchell, D. Grace and T. Clarke, “Use of Q-Learning approaches for practical medium access control in wireless sensor networks,” Engineering Applications of Artificial Intelligence, 55, pp. 146-154, 2016. [5] Y. Chu, S. Kosunalp, P.D. Mitchell, D. Grace and T. Clarke, “Application of reinforcement learning to medium access control for wireless sensor networks,” Engineering Applications of Artificial Intelligence, 46, pp. 23-32, 2015. [6] I. Demirkol, C. Ersoy and F. Alagoz, F, “MAC Protocols for Wireless Sensor Networks: a Survey,” Communications Magazine, 2006, 44, pp. 115–121, 2006. [7] M. A. Yigitel, O. D. Incel and C. Ersoy, “QoS-aware mac protocols for wireless sensor networks: A survey,” Computer Networks, 55, pp. 1982-2004, 2011. [8] P. Huang, L. Xiao, S. Soltani, M.W. Mutka and N. Xi, “The evolution of MAC protocols in wireless sensor networks: a survey,” IEEE Commun. Surv. Tut., 15, pp. 101-120, 2013. [9] AG. Olabi, “Energy quadrilemma and the future of renewable energy,” Energy, 108, pp. 1-6, 2016.
ACCEPTED MANUSCRIPT [10]
A. Foley, AG. Olabi, “Renewable energy technology developments, trends and policy
implications that can underpin the drive for global climate change,” Renewable and Sustainable Energy Reviews, 68, pp. 1112 - 1114, 2017 [11]
S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes: survey and
implications,” IEEE Commun. Surv. Tut., 13, pp. 443-461, 2011. [12]
F. Yang, L. Du, W. Chen, J. Li, Y. Wang and D. Wang, “Hybrid energy harvesting
for condition monitoring sensors in power grids,” Energy, 118, pp. 435-445, 2017. [13]
S. Kosunalp, “MAC protocols for energy harvesting wireless sensor networks:
survey,” ETRI Journal, 17, pp. 804-812, 2015. [14]
M.Y. Naderi, P. Nintanavongsa and K.R. Chowdhury, “RF-MAC: a medium access
control protocol for re-chargeble sensor networks powered by wireless energy harvesting,” IEEE Transactions on Wireless Communications, vol. 13, no. 7, pp. 39263937, 2014 [15]
S. Kosunalp and MB Tabakcioglu, “MAC protocols for RF energy-harvesting
wireless sensor networks: a survey,” International Journal of Computer Science and Information Security, 15(4), pp. 28-34, 2017. [16]
A. Kansal, J. Hsu, S. Zahedi and M.B. Srivastava, “Power management in energy
hasrvesting sensor networks,” ACM Transactions on Embedded Computing Systems, 6, Article 32, 2007. [17]
D.K. Noh & K. Kang, “Balanced energy allocation scheme for a solar-powered sensor
system and its effects on network-wide performance,” Journal of Computer and System Sciences, 77, pp. 917-932, 2011. [18]
J.C. Piorno, C. Bergonzini, D. Atienza & T.S. Rosing, “Prediction and management in
energy harvested wireless sensor nodes,” In Proc. IEEE Wireless VITAE’09, 2009, pp. 610. [19]
A. Cammarano, C. Petrioli & D.Spenza, “Online energy harvesting prediction in
environmentally powered wireless sensor networks,” IEEE Sensors Journal, 16, pp. 6793-6804, 2016. [20]
S. Kosunalp, “A new energy prediction algorithm for energy-harvesting wireless
sensor networks with Q-learning,” IEEE Access Journal, 4, pp. 5755-5763, 2016.
ACCEPTED MANUSCRIPT [21]
S. Kosunalp, “A performance evaluation of solar energy prediction approaches for
energy-harvesting wireless sensor networks,” International Journal of Applied Mathematics, Electronics and Computers, 4, pp. 424-427, 2016. [22]
http://www.ritm.gov.tr, last accessed on 28.12.2016.
[23]
NREL: Measurement and Instrumentation Data Center. [Online]. Available:
http://www.nrel.gov/midc/, , last accessed on 28.12.2016.
Figure 1. Architecture of a sensor node powered by a small battery.
Figure 2. EH sensor node architecture with real hardware.
ACCEPTED MANUSCRIPT
Figure 3. An example scenario of EWMA, α = 0.5, α = 0.3 and α = 0.7.
Figure 4. Energy prediction architecture of WEP.
ACCEPTED MANUSCRIPT 3000
Actual Harvested Energy EWMA Pro-Energy WEP
2500
Energy [W]
2000
1500
1000
500
0 6480
6500
6520
6540
6560
6580
6600
6620
6640
Time, Slots
(a) 4500
Actual Harvested Energy EWMA Pro-Energy WEP
4000 3500
Energy [W]
3000 2500 2000 1500 1000 500 0
7160
7170
7180
7190
7200
7210
7220
Time, Slots
(b) Figure 5. Prediction accuracy for three schemes (a) RITM, (b) NREL.
ACCEPTED MANUSCRIPT TABLE I Prediction error ratios for three schemes in RITM, α = 0.5. EWMA
Pro-Energy
WEP
0.559
0.179
0.086
TABLE II Prediction error ratios for three schemes in NREL, α = 0.5. EWMA 1.25
Pro-Energy
WEP
0.755
0.436
ACCEPTED MANUSCRIPT
A new energy prediction model is proposed to supply unlimited energy for nodes
The idea relies on utilization of most recent information about energy generation
Results highlight the efficiency of the scheme outperforming the existing approaches
Selahattin Kosunalp PII:
S0360-5442(17)30965-9
DOI:
10.1016/j.energy.2017.05.175
Reference:
EGY 10983
To appear in:
Energy
Received Date:
30 December 2016
Revised Date:
24 May 2017
Accepted Date:
28 May 2017
Please cite this article as: Selahattin Kosunalp, An Energy Prediction Algorithm for Wind-Powered Wireless Sensor Networks with Energy Harvesting, Energy (2017), doi: 10.1016/j.energy. 2017.05.175
This is a PDF file of an unedited manuscript that has been accepted for publication. As a service to our customers we are providing this early version of the manuscript. The manuscript will undergo copyediting, typesetting, and review of the resulting proof before it is published in its final form. Please note that during the production process errors may be discovered which could affect the content, and all legal disclaimers that apply to the journal pertain.
ACCEPTED MANUSCRIPT
An Energy Prediction Algorithm for Wind-Powered Wireless Sensor Networks with Energy Harvesting Selahattin Kosunalp 1Department
of Industrial Engineering, Faculty of Engineering, University of Bayburt, 69000 Bayburt, Turkey 2Central Research Laboratory, University of Bayburt, 69000 Bayburt,
Email: [email protected]
Abstract— Energy harvesting (EH) from environmental energy sources has the potential to ensure unlimited, uncontrollable and unreliable energy for wireless sensor networks (WSNs), bringing a need to predict future energy availability for the effective utilization of the harvested energy. The majority of previous prediction approaches have exploited the diurnal cycle dividing the whole day into equal-length time slots in which predictions were carried out in each slot independently. This is not, however, efficient for wind energy as it exhibits non-controllable behaviour in that the amount of energy to be harvested varies over time. This paper proposes a novel approach to predict the wind energy for EH-WSNs depending on the energy generation profile of latest condition. The distinctive feature of the proposed approach is to consider the recent conditions in current-day, instead of past-day’s energy generation profiles. The performance of the proposed algorithm is evaluated using real measurements in comparison with state-of-art approaches. Results show that the proposed strategy significantly outperforms the two popular energy predictors, EWMA and ProEnergy. Index Terms— Energy harvesting, wind power, wireless sensor networks, energy prediction.
1. Introduction The main body of wireless sensor networks (WSNs) includes a collection of resourceconstrained sensor nodes performing a common task [1]. A typical sensor node senses an environmental parameter, such as temperature or movement, which is then processed to be transmitted to a central point, called sink or base-station. Sensor nodes are solely powered by limited small batteries as depicted in Fig. 1 with all components. Traditionally, sensor nodes are randomly deployed in harsh and remote areas where a physical access to the location of the sensor nodes is often impractical resulting in recharging or replacement of the battery a difficult task. The energy constraints of sensor nodes may limit the functions of the WSNs. A
ACCEPTED MANUSCRIPT successful deployment of long-lasting WSNs relies on the effective utilization of the current energy level. The emphasis has therefore been placed on maximizing the operation lifetime of WSNs. Wireless communication consumes the majority of energy source when compared to energy consumptions in sensing and processing parts. Therefore, an intelligent coordination of channel capacity among sensor nodes is required to effectively manage the energy [2]. It is well-understood that energy is wasted through idle-listening, overhearing, extra control packet exchange and collisions [3]. Medium access control (MAC) protocols are necessarily designed to minimize aforementioned energy wastage reasons through assigning intelligent capacity of channel and coordinating communication in collision-free manner [4, 5]. A huge number of MAC protocols have already been proposed for WSNs with the core responsibility of energy-efficiency [6-8]. However, the energy will eventually deplete resulting in nodes losing their operation. In order to handle the bottleneck of energy in all energy-limited systems, renewable energy sources in the environment can be exploited as an unlimited energy source [9, 10]. General principle of this idea is to convert an existing ambient energy into electricity supplying the energy burden of the systems. Therefore, EH process from the surrounding environment would potentially provide an infinite energy source through continuous energy harvesting. Sensor nodes with the capability of EH exploit the ambient energy by employing an EH unit [11] as presented in Fig. 2. This unit includes an energy harvester, such as a solar panel or a wind turbine, and an energy storage component to accumulate the harvested energy. As energy is infinite, sensor nodes can operate perpetually, subject to hardware failure. Therefore, traditional MAC protocols may not be applicable to WSNs with EH (EH-WSN). This paradigm opens a new forum to the researchers. The fundamental design criterion has shifted from energy-efficiency to maximizing utilization of the energy harvested [12]. A number of MAC protocols explicitly designed for EH-WSNs have been surveyed in [13]. The proposed protocols considers the current rate of energy harvesting and the amount of energy stored. Another way of powering sensor nodes is RF energy transfer which relies on broadcasting energy packets by an energy emitting node [14]. In most cases, the energy packets are periodically sent to avoid instant energy shortages. However, periodic transfer of same level of energy would not be efficient as energy requirement of each node may exhibits a varying-characteristic. One of the major issues in RF energy transfer is the use of same frequency for both data and energy transfer. This brings necessity to carefully arrange the
ACCEPTED MANUSCRIPT sessions of data and energy. Also, the number of energy emitting nodes should be appropriately adjusted to particularly cover large areas. In order to solve these problems, new MAC protocols have been gaining much attention [15]. Solar and wind energy are the most effective energy source because of their high power intensity. The amount of energy to be harvested is time-varying depending on the environmental conditions. An accurate prediction of future energy availability is therefore of paramount importance. Most of the proposed energy prediction algorithms take solar energy into consideration. This is because the prediction of solar energy is relatively straightforward when compared to other forms of ambient energy. Existing approaches exploits the diurnal cycle of sun and assumes that the amount of energy harvested in a particular time slot of a day is similar to the energy harvested in the same time slot of the following days [16]. Basically, these approaches divides a day into equal-length time slots and predicts the energy independently in each slot. Due to the significant prediction errors when the environmental or weather conditions change frequently, a number of prediction approaches considers the current conditions in order to minimize the prediction errors [16-21] and these schemes are introduced in detail in the following section. Sensor nodes are often required to perpetually operate and an energy shortage would interrupt the operation of a typical WSN. The level of solar energy to be harvested in a particular time depends highly on intensity of sun. It is obvious that there will be no energy to harvest during night-time. The time of darkness takes long hours and sensor nodes must use their accumulated energy. A sensor node would potentially deplete its energy in night-time resulting in losing its operation. Therefore, the node becomes useless until energy harvesting begins, so that some important information may be lost. The main advantage of wind energy is its existence for a whole day. This will give sensor nodes more chances to stay alive in nigh-time. Nevertheless, prediction of wind energy is an important issue to establish. The trend of wind energy is more likely to change from day to day. Unlike solar energy, the predictions do not necessarily consider the diurnal cycle for an accurate estimation of wind energy. It is believed that recent information about the wind energy would carry more important information about the near-future energy prediction. This paper presents a new algorithm that predicts the future energy availability for wind energy. The performance of the proposed algorithm is compared with state-of-art approaches. In order to reflect the actual performance of the approaches, we obtained real measurements from two different sources which cover the period of 1-year in 2015. Section II presents the
ACCEPTED MANUSCRIPT description of state-of-art approaches covered in this paper and the fundamentals of the proposed algorithm. Performance outputs of the proposed approach in comparison to other approaches are presented in Section III. Section IV concludes the paper.
2. Existing Energy Prediction Approaches This section presents the details of the energy prediction algorithms which are compared with the proposed algorithm. The first scheme, called exponentially-weighted moving average (EWMA), can be accepted as a baseline scheme as many prediction algorithms have used the theme of EWMA [16]. It introduces the concept of equal-length time slots in which the energy prediction is explicitly performed in each slot. Weather-conditioned moving average (WCMA) improves the performance of EWMA considering not only the past day observation, but also the conditions in the present day [18]. PROfile Energy (Pro-Energy) targets at forecasting both solar and wind power [19]. Similar to EWMA, a day is considered as a collection of time slots. Then, the proposed algorithm is described in details.
2.1 Exponentially-weighted moving average (EWMA) The basic idea of EWMA, as discussed previously, is to split the day into identical time slots in terms of slot duration [16]. It assumes that the energy generation pattern in consecutive days exhibits a similar behaviour. Therefore, EWMA uses historical information of the energy generation pattern as presented in Equation (1). For this purpose, the last amount of harvested energy (H) and estimated energy (E) by EWMA are summed with a weighting factor, 0<α<1, arranging the importance of the H and E. A high value of α corresponds to less importance of the last-harvested energy and vice versa. In this equation, d represents the present day and n is the slot identifier. E(d, n) = α E(d-1, n) + (1 – α) H(d-1, n)
(1)
A distinctive feature of EWMA is its adaptation to seasonal variations through a weighted average of the harvestable energy in each slot in association with a set of energy available in past days. In order to mitigate such seasonal change, the value of α should be accurately arranged. In the original paper of EWMA, α value of 0.5 has been experimentally validated as an optimum value minimizing the prediction errors. A typical example of EWMA with 24 slots in adapting to a seasonal change using three different α values is presented in Fig. 3. This example shows the behaviour of EWMA process with various α values in a particular slot when the amount of harvested energy suddenly becomes four times greater than the expected energy. It is assumed that the harvested energy continues to be the same value. Due
ACCEPTED MANUSCRIPT to the higher real energy measurement, smaller α values cause the prediction to converge to the real energy measurement more aggressively. Therefore, the choice of α value becomes a crucial issue to establish which impacts on the time efficiency of the adaptation to seasonal changes.
2.2 Weather-conditioned moving average (WCMA) WCMA has it basics on EWMA with the introduction of current weather conditions in order to account for the influence of weather conditions of the current day [17]. This improvement essentially reduces the deficiencies of the EWMA in particularly frequently changing weather conditions. WCMA maintains a collection of past days energy values in a matrix E(i, j), where j indicates a particular slot on the ith day. When estimating an energy value in a slot, WCMA collects the amount of energy harvested in the previous slot, the average energy value of the relevant slot for a pre-defined number of previous days and a weighting factor representing the current solar conditions: (2)
E(d, n) = α H + (1 – α) MD(d, n)GAP
Here, α is the weighting factor as defined in EWMA. H carries the last amount of energy harvested and MD is the average of the sample for D previous days. The main contribution in this scheme is the GAP value to measure the current solar conditions. A mean value of a sample (n) on the dth day is calculated for D past days as: d−1
M(d, n) =
�
E(𝑖, 𝑛)
𝑖=𝑑−D
(3)
D General principle of GAP value is to observe the trend of the previous slots with respect to the previous days. To do this, the behaviour of solar energy outputs in K number of previous slots is considered. In order to calculate the GAP value, a vector with K elements, V = [V1, V2,…, VK], is initially defined to contain a numerical representation of energy generation profile in each K past slots. Each value of V is actually the ratio of the energy harvested to the average value: Vk = E(d, n-K+k) M(d, n-K+k)
(4)
ACCEPTED MANUSCRIPT It is obvious that the information in the recent slots would carry more important news about the current status of weather. WCMA assigns more importance to the recent slots through an another vector, P = [P1, P2, …, PK], where each element is defined according to the distance to the actual slot: Pk = k K
(5)
Once the values of the V vector are obtained, the GAP value is calculated as: GAP = V.P
(6)
∑P
2.3 Profile energy prediction model (Pro-Energy) PROfile energy (Pro-Energy) also discretizes each day into time slots and energy prediction for each time slot is derived at the beginning of the associated time slot [18]. Similar to WCMA, Pro-Energy utilizes the amount of energy harvested in the last slot. The main concept is to maintain the energy generation profiles of D past days in a matrix, E(i, j). This matrix is then used to find out the most similar day to the current day in terms of energy generation similarity. Therefore, Pro-Energy combines the last energy harvested and the most similar energy profile from the recent past as presented in equation 7. Ê(d, n) = α H + (1 – α) EMS
(7)
Where H is the last energy harvested and Ems represents the amount of energy harvested in slot n of the most similar day. Again, α is the weighting factor as in EWMA and WCMA. A critical issue is the determination of the most similar day. The mean absolute error (MAE) of the D previous days is compared with the current day to find the most similar day, in which the day with the minimum MAE is pointed as the most similar day. The stored profile is updated with the end of a day. One typical way is to directly add the ending-day to the E matrix with the deletion of the farthest day. This is however not efficient in some cases, such as two similar days would be unnecessarily stored. Therefore, Pro-Energy performs two criterions for the decision of updating the E matrix with the current day: (1) a stored profile can stay in the matrix for a period of A days. The current day can be replaced with a day that has stayed more than A days and (2) if two or more similar days are detected, the most similar day to the current day is selected to be replaced.
ACCEPTED MANUSCRIPT For the sake of further improving the accuracy of the predictions, Pro-Energy proposes to collect the energy values from multiple profiles reducing the potential poor performance caused by a single profile. Each previous day contributes to predict the energy value with a weight factor depending on the MAE. This changes the term of the most similar day (Ems) with the weighted profile (WP). This is then computed to replace with Ems in equation 7. P
WP
w j 0
j
Esj
P 1
(8)
Where Esj represent each previous day based on their MAE and wj is computed as: wj 1
MAE Esj , C P
MAE E j 1
sj
,C
(9)
As a result, the final energy prediction equation with the multiple profile is given as: Ê(d, n) = α H + (1 – α) WP
(10)
2.4 Proposed Algorithm It may be noted that a constant value of α is the main drawback of the EWMA and WCMA resulting in a high vulnerability to frequently changing conditions. EWMA can potentially produce highly inaccurate predictions following a series of sunny and cloudy days. WCMA takes the current conditions into consideration in order to handle the main drawback of EWMA and provides an average prediction error ratio of about 20% better than EWMA. However, WCMA considers the value of energy obtained in the previous slot which leads to high prediction errors particularly at sunrise and sunset (or when wind starts and stops). ProEnergy attempts to further improve the deficiencies of EWMA and WCMA, exploiting the information obtained in the previous slot as in WCMA and finding the most similar previous day in terms of energy harvesting profile. It has been experimentally shown that Pro-Energy outperforms EWMA and WCMA significantly. Similar to WCMA, using the energy value in the previous slot would cause high prediction errors. The process of exploring the most similar previous day for a particular slot is performed through average prediction error ratios in the previous slots of past days, in which the day with the lowest error ratio is selected as the most similar day. Therefore, Pro-Energy has no mechanism to check the current conditions and is prone to finding the wrong previous energy profile. It is believed that
ACCEPTED MANUSCRIPT predictions on a daily basis would limit the prediction accuracy especially in frequently changing conditions. In this section, we introduce a novel wind energy prediction strategy for EH-WSNs, called wind energy predictor (WEP) that focuses on historical information of the latest energy generation, instead of past day’s energy generation profiles. Unlike solar energy, wind energy can be generated at any time as long as there is wind. Estimating the wind energy is usually a more difficult task, because average wind speed at a location depends on many environmental factors (e.g. trees and buildings). It is therefore possible to observe varying average wind speeds in near parts of a location. Diurnal cycle for wind energy may not be feasible in harsh environments. Instead of splitting the day into time slots, WEP generally divides the time into equal-length slots. Then, it estimates the energy of current time slot by the means of a certain number of previous time slots as depicted in Fig. 4. The main feature of WEP is to consider a certain number (N) of energy generation of previous slots. A weighted average of the N previous slots is calculated in association with the increasing index (i). An energy prediction in a single time slot is therefore carried out as in Equation (11) below. N
∑ H(i)*i
(11)
i=1
∑i H represents the actual harvested energy which is multiplied with increasing index (i) in order to give more importance to the closer time slots as most recent slots would carry the most recent information as discussed in WCMA. In order to further reduce the prediction error, we also consider the amount of energy prediction error in each of previous slots. Our motivation is to take the prediction accuracy in the previous slots into consideration for forecasting energy in a slot. To do this, an average amount of prediction error in N previous slot is added to the Eq. 11 as formed below: N
N
∑ H(i)*i
∑ H(i) – P(i)
i=1
(12)
+ i=1
∑i
N
Here, P is the predicted value in slot i. Therefore, energy prediction in a particular slot is related to the amounts of energy harvested and prediction error in previous slots.
ACCEPTED MANUSCRIPT 3. Performance Evaluation This section presents the performance of WEP in comparison to that of two energy predictors, EWMA and Pro-Energy. In all experiments of EWMA and Pro-Energy, we set the number of time slots to 24, so that a complete day is represented by 24 time slots, each of which corresponds to 1 hour duration. We did not introduce WCMA in the performance comparisons because Pro-Energy has better performance than WCMA. Another important parameter to be established is the weighting factor (α). It is set to 0.5 as this value is suggested to provide the minimum error ratio in the original papers of two approaches. ProEnergy observes a certain number of previous days to find the most similar energy generation pattern and a certain number of previous slots to decide the most similar days. As suggested in the original paper, it is set to 10 days and 7 slots, respectively. For WEP, the N is set 5 slots as it provides the best result. In order to reflect the actual performance of the approaches, we obtained real measurements from RITM project which covers the period of 1-year in 2015. RITM project is an official project established by the General directorate of Renewable Energy, the Republic of Turkey Ministry of Energy and Natural Resources [22]. It aims to develop a wind power monitoring system across Turkey. The second real-life traces of the harvested energy was obtained from the US National Renewable Energy Laboratory (NREL) [23] which is a publicly-available dataset. The main motivation behind choosing two datasets is to test the performance of the schemes in a non-frequently and frequently changing conditions, RITM and NREL respectively. Table I and II present the average prediction error ratios for three schemes over 1-year period in RITM and NREL datasets. EWMA has the worst error ratio because it is prone to significant errors when wind conditions change frequently. EWMA has no mechanism to check the conditions in the current day, so that it produces high errors in temporary changes. Pro-Energy has a better performance than EWMA as it estimates the energy based on the latest energy generation and the days with the most similar energy generation profile. Finding the most similar profile is performed to compare the total prediction errors among the previous days. Therefore, although the most similar day have the lowest total prediction error, a number of slots may experience significantly wrong energy profile. This can cause high prediction errors. In RITM, WEP has a two-time better performance than Pro-Energy as it only detects the latest conditions. It is concluded that WEP provides the best performance
ACCEPTED MANUSCRIPT with an average error ratio of 0.086 (~%8.6). WEP is able to detect temporal changes quickly within N number of slots. WEP also allows the closer slots to have more impact on the estimation of current time slot. In NREL, however, three schemes have worse performance as the wind conditions change very often. Nevertheless, WEP produces the most successful predictions by only considering the current conditions. Pro-Energy struggles to find the most similar past day as the energy generation pattern in consecutive days may vary significantly. Therefore, the process of performing the most similar day which uses the information of past days prediction errors as discussed above may be inefficient in highly-changing conditions. In both cases, WEP improves the prediction accuracy at a significant rate. The average prediction error ratios for a specific number of points (all points in our scenario) presented in Table I and Table II are calculated as: N
Error =
𝟏 𝐍
∑ │𝟏 ‒ 𝐇𝐏│
(13)
i=1
where H represents the amount of harvested energy in the associated slot and P is the predicted energy value from the approaches. In order to depict what the three schemes predict, we present the energy prediction accuracy for a random period in Fig. 5. Here, the amount of actual harvested energy and estimations of the three schemes are shown. As discussed above, Pro-Energy produces highly inaccurate predictions in some slots due to the wrong choice of the most similar previous day. EWMA has an exponential trend as expected producing significantly inaccurate predictions. WEP usually predicts correctly with an overall error ratio.
4. Conclusions This paper has presented a novel wind energy prediction approach (WEP) for energyharvesting wireless sensor networks. Contrary to existing studies, WEP only considers the current profile of wind energy generation while others use historical information of energy generation from past days. The performance of WEP has been tested using two real-life traces of the harvested energy. Performance results proved that WEP outperforms EWMA and Pro-Energy with a significant improvement. The results show that WEP reduces the prediction error ratio to 8.6 % from 55.9 % (EWMA) and 17.9% (Pro-Energy) in the first
ACCEPTED MANUSCRIPT dataset. For the second dataset which has a more highly variable pattern than the first, all schemes have a higher prediction error ratio and WEP, Pro-Energy and EWMA have error ratios of 43.6%, 75.5% and 125% respectively. Therefore, WEP exhibits a robust mechanism for providing more accurate predictions in frequently-changing conditions. ACKNOWLEDGEMENTS We would like to thank in advance the General directorate of Renewable Energy, the Republic of Turkey Ministry of Energy and Natural Resources, and US National Renewable Energy Laboratory (NREL) for providing real-measurement outputs. This work has been supported by Bayburt University Scientific Research and Development Support Program (BAP) in Turkey. The project number is 2016/01-11. References [1] I. F. Akyildiz, W. Su, Y. Sankarasubramaniam and E. Cayirci, “A survey on sensor networks,” Communications Magazine, 40, pp. 102-114, 2002. [2] C. Gherbi, Z. Aliouat and M. Benmohammed, “An adaptive clustering approach to dynamic load balancing and energy efficiency in wireless sensor networks,” Energy, 114, pp. 647-662, 2016. [3] S. Kosunalp, P.D. Mitchell, D. Grace and T. Clarke, “Experimental study of capture effect for medium access control with ALOHA,” ETRI Journal, 37(2), pp. 359-368, 2015. [4] S. Kosunalp, Y. Chu, P.D. Mitchell, D. Grace and T. Clarke, “Use of Q-Learning approaches for practical medium access control in wireless sensor networks,” Engineering Applications of Artificial Intelligence, 55, pp. 146-154, 2016. [5] Y. Chu, S. Kosunalp, P.D. Mitchell, D. Grace and T. Clarke, “Application of reinforcement learning to medium access control for wireless sensor networks,” Engineering Applications of Artificial Intelligence, 46, pp. 23-32, 2015. [6] I. Demirkol, C. Ersoy and F. Alagoz, F, “MAC Protocols for Wireless Sensor Networks: a Survey,” Communications Magazine, 2006, 44, pp. 115–121, 2006. [7] M. A. Yigitel, O. D. Incel and C. Ersoy, “QoS-aware mac protocols for wireless sensor networks: A survey,” Computer Networks, 55, pp. 1982-2004, 2011. [8] P. Huang, L. Xiao, S. Soltani, M.W. Mutka and N. Xi, “The evolution of MAC protocols in wireless sensor networks: a survey,” IEEE Commun. Surv. Tut., 15, pp. 101-120, 2013. [9] AG. Olabi, “Energy quadrilemma and the future of renewable energy,” Energy, 108, pp. 1-6, 2016.
ACCEPTED MANUSCRIPT [10]
A. Foley, AG. Olabi, “Renewable energy technology developments, trends and policy
implications that can underpin the drive for global climate change,” Renewable and Sustainable Energy Reviews, 68, pp. 1112 - 1114, 2017 [11]
S. Sudevalayam and P. Kulkarni, “Energy harvesting sensor nodes: survey and
implications,” IEEE Commun. Surv. Tut., 13, pp. 443-461, 2011. [12]
F. Yang, L. Du, W. Chen, J. Li, Y. Wang and D. Wang, “Hybrid energy harvesting
for condition monitoring sensors in power grids,” Energy, 118, pp. 435-445, 2017. [13]
S. Kosunalp, “MAC protocols for energy harvesting wireless sensor networks:
survey,” ETRI Journal, 17, pp. 804-812, 2015. [14]
M.Y. Naderi, P. Nintanavongsa and K.R. Chowdhury, “RF-MAC: a medium access
control protocol for re-chargeble sensor networks powered by wireless energy harvesting,” IEEE Transactions on Wireless Communications, vol. 13, no. 7, pp. 39263937, 2014 [15]
S. Kosunalp and MB Tabakcioglu, “MAC protocols for RF energy-harvesting
wireless sensor networks: a survey,” International Journal of Computer Science and Information Security, 15(4), pp. 28-34, 2017. [16]
A. Kansal, J. Hsu, S. Zahedi and M.B. Srivastava, “Power management in energy
hasrvesting sensor networks,” ACM Transactions on Embedded Computing Systems, 6, Article 32, 2007. [17]
D.K. Noh & K. Kang, “Balanced energy allocation scheme for a solar-powered sensor
system and its effects on network-wide performance,” Journal of Computer and System Sciences, 77, pp. 917-932, 2011. [18]
J.C. Piorno, C. Bergonzini, D. Atienza & T.S. Rosing, “Prediction and management in
energy harvested wireless sensor nodes,” In Proc. IEEE Wireless VITAE’09, 2009, pp. 610. [19]
A. Cammarano, C. Petrioli & D.Spenza, “Online energy harvesting prediction in
environmentally powered wireless sensor networks,” IEEE Sensors Journal, 16, pp. 6793-6804, 2016. [20]
S. Kosunalp, “A new energy prediction algorithm for energy-harvesting wireless
sensor networks with Q-learning,” IEEE Access Journal, 4, pp. 5755-5763, 2016.
ACCEPTED MANUSCRIPT [21]
S. Kosunalp, “A performance evaluation of solar energy prediction approaches for
energy-harvesting wireless sensor networks,” International Journal of Applied Mathematics, Electronics and Computers, 4, pp. 424-427, 2016. [22]
http://www.ritm.gov.tr, last accessed on 28.12.2016.
[23]
NREL: Measurement and Instrumentation Data Center. [Online]. Available:
http://www.nrel.gov/midc/, , last accessed on 28.12.2016.
Figure 1. Architecture of a sensor node powered by a small battery.
Figure 2. EH sensor node architecture with real hardware.
ACCEPTED MANUSCRIPT
Figure 3. An example scenario of EWMA, α = 0.5, α = 0.3 and α = 0.7.
Figure 4. Energy prediction architecture of WEP.
ACCEPTED MANUSCRIPT 3000
Actual Harvested Energy EWMA Pro-Energy WEP
2500
Energy [W]
2000
1500
1000
500
0 6480
6500
6520
6540
6560
6580
6600
6620
6640
Time, Slots
(a) 4500
Actual Harvested Energy EWMA Pro-Energy WEP
4000 3500
Energy [W]
3000 2500 2000 1500 1000 500 0
7160
7170
7180
7190
7200
7210
7220
Time, Slots
(b) Figure 5. Prediction accuracy for three schemes (a) RITM, (b) NREL.
ACCEPTED MANUSCRIPT TABLE I Prediction error ratios for three schemes in RITM, α = 0.5. EWMA
Pro-Energy
WEP
0.559
0.179
0.086
TABLE II Prediction error ratios for three schemes in NREL, α = 0.5. EWMA 1.25
Pro-Energy
WEP
0.755
0.436
ACCEPTED MANUSCRIPT
A new energy prediction model is proposed to supply unlimited energy for nodes
The idea relies on utilization of most recent information about energy generation
Results highlight the efficiency of the scheme outperforming the existing approaches