A Hybrid Handover Forecasting Mechanism Based on Fuzzy Forecasting Model in Cellular Networks

Hua Qu, Yanpeng Zhang＊, Jihong Zhao, Gongye Ren, Weipeng Wang

1 School of Software Engineering, Xi’an Jiaotong Univ., Xi’an 710049, China

2 School of Electronics and Information Engineering, Xi’an Jiaotong Univ., Xi’an 710049, China

3 School of Communication and Information Engineering, Xi’an Univ. of Posts & Telecommunication, Xi’an 710061, China

4 Suzhou Caiyun Networking Technologies Co., Ltd., Suzhou 215000, China

I. INTRODUCTION

Due to the user’s frequent and inevitable mobility, handover management becomes one of the key technologies to guarantee the user experience. For satisfying the growth of data traffic in cellular networks, the small cell networks are proposed. The reduction of coverage of cells makes the handover decision more difficult. On the one hand, increasing interest has been observed on ways to adopt context-aware techniques [1] to improve the mobility and resource management in future wireless networks. On the other hand, in 3GPP Release 8 [2], the history mobility information, which could be recorded by base stations,was first proposed. The user’s future mobility,which could be obtained by context-aware scheme based on user’s history mobility, could help with handover management. Thus, the context-aware technique could become an effective method to provide seamless services in future wireless networks.

Originally, received signal strength (RSS)was the only index that was used to make handover decision. Then, reference signal receiving power (RSRP) was introduced into LTE to replace RSS. To develop an efficient handover mechanism, standard RSS-based mechanisms[3,4] focus on adjusting parameters to handle various scenarios. However, these handover mechanisms based on time to trigger (TTT)and hysteresis cannot achieve both timely and exact handovers. Recently, forecasting models [5,6] are introduced to solve this problem.Most of these researches only consider either short-term or long-term forecasting. However,these two kinds of forecasting are complementary, and combination of them could help with executing timely handover as well as preventing from ping-pong handover. In [7], authors present an admission control strategy which aims at optimizing traffic load by considering both short-term and long-term factors, but the long-term factors are used to deal with the load-balance problem of base station instead of handover decision.

Since the parameters used to make handover decision, such as RSRP and location,usually contain noise and measurement error,we apply fuzzy forecasting model [8], which could handle data that are ambiguous and inaccurate, to deal with them. Besides, most of the existing handover mechanism does not consider position, except in some high-speed scene, because of the indirectly relationship between trajectory and user experience. However, the trajectories are more suitable for long-term forecasting, and forecasting result of the position is relevant to the RSRP, which could be used to avoid ping-pong handover.Hence, we developed a hybrid handover forecasting mechanism, with forecast of both RSRP and position. Furthermore, the proposed mechanism could work with standard mechanism compatibly, to handle the cold-start problem.

Our research makes three main contributions: 1) we apply a fuzzy forecasting model for handling the imprecise data; 2) we design a hybrid forecasting mechanism, which contains a long-term model for trajectory and a short-term model for RSRP, and 3) we make the proposed mechanism compatible with the standard mechanism.

The rest of this paper is organized as follows. Section II introduces some related researches. Section III presents some basic concepts and algorithms. Section IV gives the detailed procedure of the proposed mechanism. Section V shows the simulation results and the conclusion is provided in Section VII.

The authors designed a hybrid handover forecasting mechanism based on fuzzy forecasting model.

II. RELATED WORK

In order to perform handover, extra resources will be consumed. A seamless and effective handover could improve the user experience.Conversely, a wrong handover would waste resources, and even lead to connection dropping. One of the most effective techniques to make reasonable handover decision is context-aware based handover mechanism, which utilizes the future condition of users to make handover decision [1].

The handover decision algorithms are classified as four categories:

1) Location based decision algorithms. This kind of algorithm provides location information, which could generate a comprehensive knowledge about the actual network and effect a better handover decision. [9] presents a target cell prediction method in LTE femtocells network via simple Markov Chain technique.

2) Mobility or speed based decision algorithms. Usually the speed, direction and trajectory of user equipment (UE) are taken into account in these algorithms. [10] provides an efficient handover mechanism in which both moving direction prediction and TTT adjustment are taken into consideration for processing to lower the unnecessary handover.

3) Policy based decision algorithms. In a policy based approach, different kinds of information could be used for optimizing the handover process. [11] designs fuzzy inference rules for handover decision, and applies fuzzy ranking technique for target base station selection. In [12], the authors propose a handover decision strategy which applies a Mamdani fuzzy logic to make decision with RSS, data rate, velocity of mobile terminal, and traffic load.

4) Learn based decision algorithms. Machine learning is applied to learn the context or to make the decision in these algorithms.[13] presents a Fuzzy Q-Learning-based ap-proach which aims at providing a generic basis for enabling self-optimizing and self-healing network operations.

III. PRELIMINARIES

Fuzzy time series model [8] is a forecasting framework that could deal with the data that are incomplete, imprecise and ambiguous. It was first proposed by Song and Chissom in 1993. This fuzzy forecasting model fuzzifies the time series by replacing the values with fuzzy sets defined on the universe of discourse[14]. Hence, the measurement errors hardly make a difference to the forecast results.

The fuzzy time series forecasting model uses a four-step framework to make forecast:1) define the universe of discourse and partition it into intervals; 2) determine the fuzzy sets on the universe of discourse by using the partition results, and fuzzify the time series into fuzzy time series; 3) build the model of the existing fuzzy logic relationships in the fuzzified time series; 4) make forecast and defuzzify the forecast values.

IV. HANDOVER FORECASTING MECHANISM

4.1 System model

In wireless network, several macrocells and microcells are deployed in the region where UEs move, and the UEs move to their destinations based on their needs. Besides, UEs usually move along the street, road and other paths.

UE can update its position everyτpseconds,referred to as UE position updating interval.We assume that each UE can get its location information periodically by using localization techniques.

In addition, each UE can obtain the RSRP of all cells every τrseconds, which may become its serving cell. The RSRP largely influences the user experience. Furthermore, weak RSRP would lead to connection dropping.

4.2 A short-term fuzzy forecasting model for RSRP

Most of the handover mechanisms consider the RSRP as an important parameter. Therefore, many researches focus on adjusting the hysteresis and TTT to fit various scenarios.However, no matter how these thresholds are defined, this simple model cannot handle all of the scenarios in the real world, it just trades off between handover failure and ping-pong handover [15].

Wrong handovers comprise ping-pong handover and handover failure, and they are mainly caused by the lack of the future information. Nowadays, the storage and computing ability of communications equipment is enhanced, which allows a complicated and accurate model for forecasting the future RSRP. Besides, since the RSRP is influenced by slow fading, shadow fading and noise, the forecasting method should have certain noise tolerance to get rid of these interferences. Fortunately, the fuzzy forecasting model can deal with this problem.

To achieve forecasting handover mechanism,there is a period of RSRP, which consists ofa sample of historical RSRP data. To forecast=ri+tp, a forecasting model is built by using the method presented in [16]. Besides,handover mechanism demands RSRP of a period with the length equaled TTT to make decision, which means that the forecasting period’s length is at most TTT, hence,0≤tp≤TTT.

The historical datar1,r2,...,rtis transferred into samples of historical RSRP time series with the length equaled tok:

These samples are clustered by Fuzzy C-means [17]. Then, cluster centers V are ob-tained. The grade of membership εi,jbetweensample and cluster centercan be calculated as defined in [17].

By adopting the algorithm in [16], we get a matrix Q which defines a weight vectorof linear combination for each cluster. Then, the forecast value ofcan be calculated as followed:

By applying this method, we could build forecasting model for each time point by changing the value oftp.

In short-term forecasting model for RSRP,since users always keep stable motion in a short period, the user’s trajectories and habits have little effect on the trend of the RSRP.Instead, transmitting power and channel model of cell largely determine the trend of the RSRP. Thus, for each cell, all of the RSRP data related are used to train its RSRP forecasting model, no matter which UE it belongs to. Finally, we get a set of RSRP forecasting models for all cells.

With RSRP forecasting models of all cells,the short-term handover forecasting mechanism could be executed as followed:

1) The UE monitors RSRPs of all cells.Once it finds a cell, whose RSRP is close to the serving cell’s, it will regard this cell as a target cell, and add it to the option list. Then,a request is immediately sent to get its RSRP forecasting model;

2) Check the option list at regular intervals τr. If the option list of target cell is empty or the serving cell’s RSRP forecasting models has not been obtained, end this process.

3) For each target cell:

3.1) If the target cell’s RSRP is much worse than the serving cell’s, remove the cell from the option list, and go to the step 4). Besides,if either of the forecasting models for RSRP of serving cell and the target cell is missing,ignore this cell and continue the processing;

3.2) Forecast the RSRP of the target cell.Suppose the target cell’s RSRP is greater than serving cell, which lastst, so the length of period to forecast can be expressed as

4) If all of the target cells in the option list have been processed, forecast the RSRP of serving cell, the length of period to forecast can be calculated as

5) Take both of the historical data and the forecasting results into account. Check all of the cells in the option list. If the RSRP of a target cell is greater than the RSRP of serving cell, and it will last for a period time longer thanTTT, the UE would make a suggestion of executing a handover to this target cell.

Generally, when a cell is added to the option list, its RSRP will be concerned for a period. During this time, the RSRP related to this cell will be forecasted for several times. Thus,transferring the forecasting model is faster and more efficient than transferring the forecasting results.

4.3 A long-term fuzzy forecasting model for trajectory

Users do not always keep still when they use their UEs. The position where the user will be plays an important role in handover decision,because almost all of the channel models regard distance as the most valuable parameter.Especially in some researches, distance is used to make horizontal handover decision. However, position is not the only independent variable in the channel model, and channel model in the real world is much more complicated than the channel model they used. Therefore,it is not enough to make handover decision only with position.

As the coverage of cell become smaller[18], the short-term forecast of the RSRP cannot avoid ping-pong handover when a user quickly passes through a small cell. Yet an accurate long-term forecasting model of trajectory could solve this problem. Moreover,the position is suit for long-term forecasting.Consequently, though the position forecasting could not help making an exact handover decision, it could prevent from ping-pong hando-ver and handover failure.

No matter which localization technique, the positioning error [19] is inevitable because of multipath effect, atmosphere and other factors.Generally, several localization techniques,are applied by UE, such as Time of Arrival(TOA), Enhanced Observed Time Difference(E-OTD), Assisted Global Positioning System(A-GPS). The positioning errors vary from 5m to 100m. Hence, the forecasting models should handle this error. In our research, a fuzzy forecasting model is applied for dealing with this problem. In the fuzzy forecasting model, positions are fuzzified into fuzzy sets,and most of the positioning errors are ignored.

In order to fuzzify the trajectories, a set of fuzzy sets need to be designed. Details of the construction of fuzzy sets will be described as follows:

1) Define the discourse of universe of trajectories. The universe can be represented by a rectangular area, which covers all of the position points in the trajectory. For a trajectory consist oftposition points,ith point of this trajectory can be presented aspi=(xi,yi) in Cartesian coordinates, and the scope of this trajectory is a rectangle area defined by the maximums and the minimums ofxiandyi. Thus, the discourse of universeas shown in figure 1, can be calculated by expending this rectangle area.

Fig. 1. The partition of discourse of universe of trajectory and the related fuzzy sets.

2) To fuzzify trajectories, the discourse of universe is divided into several regions by grid mesh. Serval partition points are generated in both X-axis and Y-axis. With these points, the grid mesh is built. As shown in figure 1, these regions have different sizes, and each region corresponds to a unique fuzzy set.

3) With these fuzzy sets, trajectories could be fuzzified by replacing positions with fuzzy sets. For each point in the trajectory, the point is fuzzify into fuzzy setSionly when this point belongs to the region corresponding toSi.

4) Evaluate these fuzzy sets by using the forecasting root mean square error (RMSE)when experiment on the testing sets;

5) Find an available set of fuzzy sets by using heuristic optimization algorithm, which means finding a set of suitable partition points.There are plenty of works focus on finding the valuable partition points, here we adopted a heuristic optimization algorithm described in[20].

Since this model treats different positions in the same region as the same fuzzy set, most of the positioning errors are eliminated after preprocessing the origin data, once a valuable partition was found. Thus, most of the influence from the positioning error can be avoided.

Because each user has its own pattern of life, the model of trajectory varying according to the user. Hence, a unique model of trajectory is built for each user. During constructing the model for a user, training set only contains trajectories that belong to the user.

The fuzzy time series forecasting model builds a set of fuzzy logic rules, and traditional models make forecasting with the weighted sum of these fuzzy logic rules. In fact, these fuzzy logic rules are cluster centers of trajectories, so each rule can be regard as a typically trajectory. Consequently, the result of our trajectory forecasting model is not a position,instead it gives, which is a set of potential positions with their probabilities.Moreover, each potential position is given by a typically trajectory obtained from a fuzzy logic rule, and the potential position is the destination of the trajectory.

The forecasting results ofcan be represented asnpis the number of cluster obtained. Then,the forecasting error can be calculated as:

4.4 Hybrid handover forecasting mechanism

As the figure 2 illustrates, unlike other handover forecasting mechanisms, both RSRP forecasting and position forecasting are introduced to make the handover decision in the proposed mechanism. On the one hand, the short-term forecasting model could make handover decision in time, but it is likely to lead to pingpong handover; on the other hand, the longterm forecasting model could avoid ping-pong handover, but its sampling interval is too large to make a handover decision. Thus, the basic handover mechanism, in combination with short-term and long-term forecasting, could get rid of its own demerits.

4.4.1 Architecture of distributed handover forecasting mechanism

After 3GPP Release 8, the base station is designed to be able to store information and perform calculation. Because of the limitation of battery and computing power, it is impracticable to make UE train all models on its own.Hence, we proposed a distributed handover forecasting mechanism.

The short-term forecasting models vary from cell to cell, as the explanation in section 4.2. Therefore, it is a waste of storage and computing power that each UE maintains a set of models corresponding to all cells. Thus,the short-term forecasting models should be stored in base stations, so that the models can easily be updated. In order to train the shortterm model, the data of the RSRP must also be stored in the base station. However, the RSRP data are collected by UEs all the time. The cost of transmission is unacceptable if all of the RSRP data are recorded and immediately transferred to base stations. Thus, we designed a lazy transmission mechanism to solve this problem, and the details are presented as follows:

Fig. 2. The architecture of the hybrid handover forecasting mechanism.

1) Once the UE finds a target cell whose RSRP is close to the serving cell’s, the UE start recording RSRP of target and serving cell;

2) The UE will stop recording immediately after encountering a ping-pong handover or a connection dropping. Besides, if the UE does not execute any handover over a period longer thanTTT, the recording would also be stopped;

3) The UE would transfer data in idle time,which means the UE doesn’t have any other service to process. And each record will be transferred to the base station it is related to.

Once a UE sends a request for RSRP forecasting model, the base station transfers the short-term forecasting model to the UE. UEs keep several models, and applied Least-Recently-Used algorithm [21] for managing storage space of these models

The long-term forecasting models are built according to UEs. The UE’s trajectories are collected and stored locally, and the forecasting model is also trained at the UE.

4.4.2 Procedure of hybrid handover forecasting

With both of the short-term forecasting model for the RSRP and the long-term forecasting model for the trajectory, we designed a hybrid handover forecasting mechanism. Figure 3 describes the procedure of the proposed hybrid handover forecasting mechanism

The values of RSRP are updated in every τrseconds. Whenever there is a handover need to be decided, the forecast values of all time points, which are needed by short-term forecasting handover mechanism, should be updated by using the latest data collected.As for position, at a given time point, there are only one forecasting value of a specific time point need to be forecasted. During the detecting period of position, this forecasting value wouldn’t change since the historical data is unchanged. Once the historical data of position changes, this forecasting value is no longer need in our mechanism.

Short-term forecasting handover suggestion is generated by the method described in section 4.2. At the same time, the UE makes forecast of position with long-term model for trajectory. Whether the handover decision will be performed is depending on the forecasting position.

Unlike the forecasting model of RSRP, the accuracy of the trajectory forecasting model is extremely important. Because the forecasting position is a critical step to decide whether to performed the handover suggestion and forecasting results with large prediction errors probably lead to wrong choices. Since the prediction error is inevitable, it’s necessary to detect the forecasting results with large prediction errors and discard them, which ensures that the prediction error of position hardly interferes our mechanism.

With the mobility history information, we could make forecast of the UE’s next position by using the model described in section 4.3.For each potential positionpipotenialof the longterm position model’s forecasting results, with the given handover suggestion to cellC, the UE makes a judgmentjias followed:

wherepcdenotes the base station’s position of cellC, anddcrepresents the diameter of cellC. Only when the potential position is in the coverage of the target cell, the judgement will decide to execute the handover suggestion.

Since the detecting period of position is much longer than the threshold which is used to judge ping-pong handover, if the forecasting results indicate that the UE will stay in the coverage of cell, it’s very likely UE would stay in this cell for a time that are long enough to avoid ping-pong handover; Conversely,UE will probably leave this cell soon, which means that this handover suggestion will cause ping-pong handover.

After all of the judgments corresponding to potential positions are made, the probabilities ρhof executing the handover can be calculated as:

If ρhis higher than γ, a threshold that is used to guarantee the confidence of the forecasting handover results, the handover suggestion with the largest ρhwill be performed as a handover decision. Otherwise, the decision will be ignored. Thus, the handover will be executed only when the forecasting model are firmly believed that the ping-pong handover will not happen.

4.4.3 Cooperation with standard handover mechanism

The hybrid handover forecasting mechanism will certainly improve the user experience when the forecasting results are quite accurate.To guarantee this, the forecasting of trajectory is applied to avoid wrong handover caused by forecast error of the RSRP. Moreover, a strict restriction γ is applied to the judgment made by forecast of position, to ensure that the fore-cast of trajectory is adopted only if it is highly likely accurate.

Both of the short-term and long-term forecasting model are off line, which means they don’t update immediately when new data comes. In fact, these models need not to be retrained for a long period of time, before they may be out-of-date. Every time when they are ready to be retrained, the new data collected are used as testing data to determine whether this model is trained well or not. If the RMSE of forecasting results of a model is small enough (e.g. 2dB for RSRP model, 5m for position model), the model is trained well and can be used in our mechanism.

Though the restriction ensures that the handover forecasting mechanism seldom makes wrong handover decisions, it could not take place of the standard mechanism.First, the forecasting mechanism would do nothing when it cannot make an accurate prediction, even when the handover needs to be performed absolutely. Second, sometimes the UE may not get the models it needs and the forecasting handover mechanism cannot work. Last, the forecasting mechanism needs a period time to train its models before it works,which is called cold-start. Hence, the handover forecasting mechanism should be able to work together with the standard mechanism.

To make the handover forecasting mechanism compatible with the standard mechanism,both of the standard and short-term forecasting mechanisms will work and generate handover suggestions at the same time. There are two kinds of situation after making handover suggestions:

1) The handover suggestion, made by handover forecasting mechanism, is determined to execute. No matter what happened, the UE will execute this decision.

2) The handover forecasting mechanism makes none executable decision. Instead, the standard mechanism makes one. Then, make a position prediction by long-term fuzzy forecasting model for trajectory. Only if the trajectory prediction model has been trained well, and the forecasting results decide not to execute this handover, ignore the handover decision; otherwise, execute it.

Fig. 3. The proposed hybrid handover forecasting mechanism.

V. SIMULATION RESULTS

To demonstrate the performance of the pro-posed mechanism, we make experiments on real trajectories, which was collected in the State Fair of South California. To train the fuzzy forecasting models, one-tenth of the trajectories are used as training set and others as testing set. Moreover, we applied cross-validation for experiments. We divided the whole dataset into ten parts, and we used one part,which hasn’t been tested, as testing set and nine parts as training set each time, until all of the parts have been tested.

We consider a scenario which contains both of macrocells and microcells [7]. The main related parameters used in performed simulations are summarized in Table 1.

To demonstrate the performance of the proposed mechanism, comparisons with the other handover mechanisms were made. The user experience is mainly influenced by connection dropping, handover failure and pingpong handover. The connection dropping (CD)ratiorcd, handover failure (HOF) ratiorhfand ping-pong handover (PPHO) ratiorpphoare expressed as:

whereNcdis the number of connection dropping,Nis the total number of connection,Hhfis the number of handover failures andHpphois the number of ping-pong handovers,Histhe total number of handovers. Furthermore,the overall performance indicator (OPI), which was defined in [13], is applied for evaluating the overall performance, which could be calculated as followed:

Table I. Mainly system parameters for simulations.

To compare our results with other mechanisms, we consider a location based approach as defined in [9], a policy based approach [12],a mobility based approach as defined in [4]and a learn based approach as defined in [13].

First, an experiment was performed to find out the influence on the results from TTT and hysteresis. Only the location based approach and the policy based approach mentioned were used to compared with our mechanism,because the results of the mobility based approach and the learn based approach implemented are independent of TTT and hysteresis.

Figure 4 depicts the CD ratios, the HOF ratios, the PPHO ratios and the OPIs for and the proposed mechanism with different TTTs and fixed hysteresis. The location based approach cause much higher CD ratio and HOF ratio than other two approaches, because that the trajectory dataset we used contains positioning error, which is a regular condition in the real world.

Though the location based approach makes a good performance of reducing the pingpong handover with a small TTT, the overall performance, which is indicated by OPI, is much worse than other two approaches. The experiment shows that our mechanism makes a better OPI than policy based approaches,and the main reason is that our mechanism can significant avoid connection dropping and handover failure, given a large TTT, without causing ping-pong handover.

Figure. 5 presents the CD ratios, the HOF ratios, the PPHO ratios and the OPIs with different hysteresisesand fixed TTT, as the way of figure 4. The results prove that the influences on the performance of all three approaches from hysteresis are similar, since all of these approaches are based on the standard mechanism. Besides, the proposed mechanism shows a great ability to prevent from connection dropping and handover failure, which makes it the best approaches with the highest OPI among the three approaches.

Fig. 4. The performance of different TTTs with hysteresis=1dB.

Fig. 5. The performance of different hysteresises with TTT=320ms.

We also performed an experiment to study the performance of the proposed mechanism in different scene. The dataset of trajectory was divided into two parts. One of them contains the trajectories with the average speeds lower than 20km/h, which was regards as dataset of low speed scene. The other contains the trajectories with the average speeds higher than 20km/h, which was regards as dataset of high speed scene.

Fig. 6. The performance of different approaches in low speed scene.

Fig. 7. The performance of different approaches in high speed scene.

Figure 6 shows the performance of all of the approaches we implemented, testing on the dataset of low speed scene. As former experiment results suggest, the TTT=320ms and hysteresis=2dB were set for the proposed approach. We adjusted parameters in all approaches by experiments to make them perform best on this dataset. The proposed mechanism gets the best CD ratio, HOF ratio and OPI among all approaches. Though the proposed mechanism causes a little higher PPHO ratio than the other approaches, it significantly decreases the CD ratio. As for the learn based approach, it caused too much CD which makes an awful overall performance,though it has the lowest PPHO ratio.

Another experiment on the dataset of high speed scene was performed to compare our approach with others. Figure 7 illustrates the performance of all of the approaches, with the TTT=320ms and hysteresis=2dB. The results are similar with the former experiment. All of OPIs become higher, which means a worse performance. The result suggests that the proposed mechanism is suitable for the high speed scene, which makes the lowest CD ratio and OPI, though it makes a little more PPHO than other approaches.

Besides, an experiment was performed to find out the relationship between the predict accuracy and the performance of the proposed approach. We performed the proposed mechanism with TTT=320ms and hysteresis=0dB on the whole dataset, and statistically analyze the results according to the prediction error for RSRP or position.

Figure 8 describes the relationship between the performance of the proposed mechanism and the prediction error of the RSRP. The CD and the HOF hardly occur when the prediction error is below 1 dB, which means that CD ratio and the HOF ratio mainly depend on the predict accuracy of the RSRP in our mechanism. However, the PPHO ratio with the prediction error belonged to [0.2,5) is lower than the ratio with the prediction error belong to[0,0.2), which means that the exact prediction of the RSRP could not prevent PPHO from happening.

Figure 9 presents the performance of the proposed mechanism with different prediction errors of the position. From the results,it seems that the prediction error is irrelevant to the CD ratio and HOF ratio, because that the proposed mechanism makes better CD ratio and HOF ratio when prediction is belong to [4,8) than [2,4). Overall, the PPHO ratio appeared to grow with the increase of the prediction error, but when the prediction error becomes bigger than 8m, the PPHO ratio declines. The reason is that our mechanism could recognize the sample with large prediction error, as described in section 4.4.2, and discard the prediction result. Furthermore, the accuracy of the recognition increases with the prediction error. The result shows that an accurate forecast of trajectory could improve the PPHO ratio.

At last, we performed an experiment to find the relationship between γ and the performance of the proposed mechanism. Figure 10 shows the performance with different values of γ. Obviously, the mechanism lead to connection dropping and handover failure when γ is larger than 0.8. And the ratio of pingpong handover declines with the increase of γ. Large γ means strict restriction of executing handover, which may cause c connection dropping and handover failure. Conversely,small γ may lead to ping-pong handover, because it’s more likely to execute the handover suggestion. As the result shows, the proposed mechanism performs best when γ=0.8.

VI. CONCLUSIONS

We designed a hybrid handover forecasting mechanism based on fuzzy forecasting model.The fuzzy forecasting model adopted could deal with the imprecise data. Thus, it could make reliable forecast of RSS and trajectory,which usually contain measurement errors.Besides, a hybrid handover forecasting mechanism, combined with long-term and short-term forecasting, were developed, to ensure timely and correctly handover execution. Lastly, the proposed mechanism is designed to work compatibility with standard handover mechanism,so it could guarantee the user experience when the forecasting mechanism could not work.

In future work, we consider investigating improved fuzzy forecasting model to reduce the consumption of resource. In addition, we will try to build some common models to handle the problem of cold-start.

Fig. 8. The performance of the proposed mechanism with different prediction errors of the RSRP with TTT=320s, hysteresis=0dB.

Fig. 9. The performance of the proposed mechanism with different prediction errors of the position with TTT=320ms, hysteresis=0dB.

Fig. 10. The performance of the proposed mechanism with different γ with TTT=320ms, hysteresis=0dB.

ACKNOWLEDGEMENTS

This work was supported in part by the National Major Project under Grant No.2018ZX030001016, the National Natural Science Foundation of China under Grant No.61371092, and the China Mobile Program of Ministry of Education under Grants No.MCM20150102.

[1] Makris, Prodromos, Skoutas, et al., “A Survey on Context-Aware Mobile and Wireless Networking: On Networking; and Computing Environments’ Integration,”INFOCOM IEEE Conference on Computer Communications Workshops.IEEE Xplore, 2010, pp. 1-6.

[2] 3GPP TS 36.423 V8.0.0 Release 8 Technical Specification Group Radio Access Network;Evolved Universal Terrestrial Radio Access Network (EUTRAN); X2 application protocol (X2AP).

[3] Ren G, Zhao J, Qu H, “A User Mobility Pattern Based Vertical Handoff Decision Algorithm in Cellular WLAN Integrated Networks,”IEEE The 2ed International Conference on Computer and Communications (ICCC), 2016, pp. 1550-1554.

[4] Xu J, Zhao Y, Zhu X, “Mobility model based handover algorithm in LTE-Advanced,”International Conference on Natural Computation, 2014, pp.230-234.

[5] Chen X, Mériaux F, Valentin S, “Predicting a user’s next cell with supervised learning based on channel states,”Signal Processing Advancesin Wireless Communications. Darmstadt. IEEE,2013, pp. 36-40.

[6] Miyim A M, Ismail M, Nordin R, et al., “Regressive Prediction Approach to Vertical Handover in Fourth Generation Wireless Networks,”Journal of ICT Research and Applications, vol. 8, no. 1,2014, pp. 31-48.

[7] Simsek M, Bennis M, Guvenc I, “Mobility management in HetNets: a learning-based perspective,”EURASIP Journal on Wireless Communications and Networking, vol. 2015, no. 1, 2015, pp.26.

[8] Song Q, Chissom B S, “Fuzzy time series and its model,”Fuzzy Sets & Systems, vol. 54, no. 3,1993, pp. 269-277.

[9] Amirrudin A, Ariffin S H S, Malik N N N A, et al.,“Mobility Prediction via Markov Model in LTE Femtocell,”International Journal of Computer Applications, vol. 18, no. 18, 2013, pp. 40-44.

[10] Chang F M, Wang H L, Hu S Y, et al., “An Effi-cient Handover Mechanism by Adopting Direction Prediction and Adaptive Time-to-Trigger in LTE Networks,”Computational Science and Its Applications, 2013, pp. 270-280.

[11] Vasudeva K, Dikmese S, Guvenc I, et al., “Fuzzy Based Game Theoretic Mobility Management for Energy Efficient Operation in HetNets,”IEEE Access, vol. PP, no. 99, 2017, pp. 1-1.

[12] Kustiawan I, Chi K, “Handoff Decision Using a Kalman Filter and Fuzzy Logic in Heterogeneous Wireless Networks,”IEEE Communications Letters, vol. 19, no. 12, 2015, pp. 2258-2261.

[13] Klein A, Kuruvatti N P, Schneider J, et al., “Fuzzy Q-Learning for Mobility Robustness Optimization in wireless networks,”GLOBECOM Workshops. IEEE, 2013, pp. 76-81.

[14] Zadeh L A, “Fuzzy sets,”Information & Control,vol. 8, no. 3, 1965, pp. 338–353.

[15] Vasudeva K, Simsek M, Perez D, et al., “Analysis of Handover Failures in Heterogeneous Networks with Fading,”IEEE Transactions on Vehicular Technology, vol. PP, no. 99, 2017, pp. 1-1.

[16] Askari S, Montazerin N, Zarandi M H F, “A clustering based forecasting algorithm for multivariable fuzzy time series using linear combinations of independent variables,”Applied Soft Computing, vol. 35, no. C, 2015, pp. 151-160.

[17] Bezdek J C, Ehrlich R, Full W, “FCM: The fuzzy c-means clustering algorithm,”Computers &Geosciences, vol. 10, no. 2, 1984, pp. 191-203.

[18] Gupta A, Jha R K, “A Survey of 5G Network: Architecture and Emerging Technologies,”Access IEEE, vol. 3, 2015, pp. 1206-1232.

[19] Mao G, Fidan B, Anderson B D O, “Wireless sensor network localization techniques,”Computer Networks, vol. 51, no. 10, 2007, pp. 2529-2553.

[20] Cai Q S, Zhang D, Wu B, et al., “A Novel Stock Forecasting Model based on Fuzzy Time Series and Genetic Algorithm,”The International Conference on Computational Science, 2013, pp.1155-1162.

[21] Podlipnig S, B?sz?rmenyi L, “A survey of Web cache replacement strategies,”Acm Computing Surveys, vol. 35, no. 4, 2003, pp. 374-398.