Radio Network Planning with Neural Networks

Transcription

1 Radio Network Planning with Neural Networks Thomas Binzer and Friedrich M. Landstorfer Institute of Radio Frequency Technology University of Stuttgart Pfaffenwaldring 47, 755 Stuttgart, Germany Abstract The increasing number of participants in modern mobile radio networks, especially with the prospect of the new CDMA systems, necessitates a more and more detailed and efficient radio network planning. The basis of network planning is always the prediction of the quality of transmission between the transmitter and the participant. In order to find the optimal location of the transmitters in a tolerable time, an automatic positioning of base stations in an urban area needs an accurate and fast propagation model. Presently, there is no satisfying combination of an accurate and fast propagation model and an algorithm which perform the positioning of base stations in large urban areas. In this paper a fast optimization algorithm for CDMA networks based on a fast and accurate propagation model is presented. For both, the coverage prediction and the optimization process neural networks are used. For the prediction of the coverage, which has to be optimized, a sophisticated backpropagation network is used whereas for the optimization process a self organizing map is applied. First results of the planning and optimization of an urban radio network based on these new algorithms are presented.. Introduction The state of the art of the planning of mobile phone networks concerning the locations of the basestation consists in manually setting the positions of the latter and then applying a model that yields the prediction of the field strength or other transmission parameters, respectively. This planning method is time consuming and requires a versed person who designes the radio network. With the increasing number of users in mobile phone radio networks and the associated increasing traffic the optimal exploitation of the resources becomes more and more important and consequently radio network planning becomes more and more demanding. For this reason the planning should be automatized in order to obtain an efficient radio network. An optimization tool for third generation mobile systems (CDMA) has been developed in [3], where the locations of the basestations are determined from an initial set of candidate sites, but no automatic positioning is included. In a planning loop, it is essential that the radio coverage execution is fast enough to allow several iterations. Here, as the propagation model for the optimization, a fast model based on a sophisticated backpropagation neural network is used which is described in [2]. An overview of this propagation model is given in the following section. In this work the purpose of the optimization is to suggest the most appropriate location, the transmitting power and the antenna patterns for the basestations with regard to coverage for a CDMA system. Self organizing maps (SOMs) perform the optimization process. This kind of network is as well applied to the optimization of the base station sites as to the antenna pattern optimization. These artifical neural networks are characterized by an unsupervised learning process which means that during the training no explicite target is given. The number of basestations is derived from a user given CDMA configuration and a user given traffic for the aea to be planned. 2. The Propagation Model The propagation model is based on dominant ray paths that describe the electromagnetic flow of the energy along streets and over the roof tops in an urban environment. Several dominant ray paths from the transmitters to the receiving points in the area are caculated. Each dominant ray path is evaluated seperatly and from the most dominating path for each point in the prediction area parameters like the length, the included bends, parameters describing the environment of the path like waveguiding structures are extracted. These parameters serve as inputs to the neural network. The artifical network has to be trained with measurements prior to the prediction. A sophisticated architecture was used which performs well in generalizing from the training patterns. A detailed description of the model can be

2 found in [] and [2]. This model is very fast due to an initial step in which the data base containing all the buildings is preprocessed. This preprocessing is applied only once before starting the optimization runs. 3. Self Organizing Maps Self organizing maps (SOMs) consist of neurons organized on a regular grid of dimension m (see figure ). Each neuron is represented by a d-dimensional weight vektor ~w (also called codebook vektor) where d is equal to the dimension of the input vectors ~ X = (X:::Xd). Adjacent neurons are linked to each other by a neighborhood relation so the neurons build a topology. Due to this neighborhood relation the weight vectors ~w will be arranged during the training process in such a way that they will represent the input vectors in space. During the training of the map, the Figure 2. Updating the best matching unit (BMU) and its neighbors towards the input sample marked with x(from [4]). where is a time variant parameter which is normally a monotonously falling function and t is the time. Here the following function is used: Figure. Example for a SOM with m = 2 (from [5]). input vectors ~ X are presented to the net. The algorithm selects the best matching unit (BMU) of the SOM and moves it towards the presented input vector which means that the weight vector of the neuron is adapted to the input vector. The best matchicng unit is e.g. determined by the Euclidean distance: ( ~ X ; ~ W )=min j k ~ X ; ~ W k () Not only the best matching unit (BMU) is adapted but also its neighborhood (see figure 2). The neighborhood is determined by the neighborhood function for which different kinds of functions may be used. Here the Gaussian Function is used: h c (t) = (t) p 2 exp (; k r 2 c ; r j k ) (2) 2 (t) 2 The overall learning process can be described by the following equation: W j (t +)=W j (t) +(t) h cj (t)(x(t) ; W j (t)) (3) (t) = T ; t T T is the overall optimization time. 4. Radio Network Planning Procedure The radio network planning is divided into the following steps: Estimation of cell count Location optimization of the basestations (BSs) Optimization of maximum transmitting powers Adaption of the antenna patterns After having determined the required BSs, respectively cells, the coverage in the area is predicted by a semi deterministic model based on a backpropagation neural network which is described in [] and [2]. The coverage is evaluated for reasons of control although the evaluation will not be taken into consideration for the optimization process with the self organizing maps (SOMs) because it is an unsupervised learning process. In one stage of the optimization the input vector ~ X is presented once to the net before the new configurations will be used for a new prediction. Since there is an interaction beween the location of the transmitter and (4)

3 the transmitted power the optimization is alternated. The number of steps is given by the user. When the optimization of location and the transmitted power has been finished, the optimization of the antenna patterns will follow. 4.. Estimation of Cell Count As each cell has a limited capacity one has to estimate the cell count for the area A to be planned first. The required basestations (BSs) can be determined by the number of users and the offered traffic per user in the microcell. Generally, the population living and working in the given area and the penetration with wireless terminals as well as special events have to be taken into account for the estimation. For simplicity a constant number of user per km 2 and erlang E per user is assumed here. Other parameters for the cell estimation are spectrum efficiency SE which is known for a given system, data rate DR, bandwith B and load L. With these parameters the cell number can be calculated from: user per cell = SE[ kbps MHzcell ] B[MHz] L[%] (5) DR[Kbps] E[erl] user and A cell number =(int)( user per cell user +) (6). km 2 In the algorithm a constraint is included so that every BS serves the same number of users Location optimization of the BS In order to evaluate the coverage of the area for the optimization it is necessary to determine the fieldstrength at each point. The optimization process is based on a fieldstrength prediction for microcells described in [] and [2]. The purpose of the optimization is to minimize the number of points with a received signal strength either lower (under-supplied) or higher (over-supplied) than an appropriatly chosen threshold. Over-supplied points increase the interference in the adjoining cells in the CDMA system and play therefore also a negative role. The BSs, calculated in section 4., are considered as neurons of the SOM with weights ~w = (x y z). By training the map with the covered points, the units should be moved in such a way that the over- and under-supplied points in the area concerened decrease. If only the coordinates of the geographical location of the uniformly distributed points j are used as input vectors X ~ j =(x j y j z j ), the units will also arrange in the same uniform way. In order to move the BSs, which are represented by the neurons, towards the under-supplied points and away from the over-supplied points there are mainly two possibilities: One can select and present under-supplied points more often to the net or one can increase the influence of the under supplied points and decrease the influence of the over supplied points by changing the learning rate (t). For the optimization of the BS location the learning rate is changed as follows: (t j) =(t) ( ; F (j) max ) (7) F j is the fieldstrength at point j and max is the maximum fieldstrength in the area. All coordinates with the appropriate learning rate (t j) are presented for the same number of times to the net. The initial positions of the BSs can be given by the user or can be set randomly Optimization of the Maximum Transmitting Power Besides the location optimization of the BS the transmitting power is also optimized. As with for the location optimization process, the BSs are represented by the neurons of the SOM and their weight vectors ~w whereas in this case not the coordinates are represented in the vector ~w but the transmitting powers. For the optimization only the overand under-supplied points in the area are considered as input vectors. For each under- or over-supplied point, the BS that is already the best server (for under-supplied points) respectively the BSs which contribute to the over-supply are determined. For each point j and the corresponding transmitter the desired transmitting power P j is calculated so that the transmitter supplies the corresponding point with the power of the threshold T H : P j = T H P t P pj (8) P pj is the actually received power at point j and P t is the actual power of the respective transmitter. All P pj form the input vector X. ~ In the case of the transmitting power optimization, the best matching unit (BMU) is not determined by the Euclidean distance. The BMU is the BS that was used for the calculation of the component j of the input vector X. ~ The weight vectors ~w of the neurons are initialized with the power that is necessary to supply a point with the threshold T H at distance d which follows from: s user per cell d = user (9) km Optimizing of the antenna patterns In micro cells, where many BSs are installed in a small area and where the wave propagation is characterized by

4 wave-guiding (which is also considered in the prediction model used), antenna patterns play a great role. With the antenna pattern one can e.g. supply a street canyon with two beams in the direction of the street without influencing other streets which are orthogonal to the desired direction. For the optimization of the antenna patterns in the horizontal plane, a SOM with the dimension m =is used. The weight vectors ~w =( G) of the neurons represent the angle and the gain G in the direction of the antenna. Each transmitting antenna is represented by its own map. An example of a trained and untrained map is given in figure 3. The filled points represent the neurons before the training and the unfilled points those of the trained map. The maps are initialized by an isotropic radiation. For the training of 5. Results In figure 4 first results of the new opimization algorithms with 6 BSs are shown. The following parameters were used: Spectrum efficency : 8 smhz kbit Bandwidth: 5MHz Data rate: 8 kbit s Traffic : 3 merland user User : 9 km 2 Load : 75% Threshold: ;dbm(34 dbv m ) Location/Power optimization steps : 6 Antenna Pattern optimization steps : 6 Cycle under-supplied over-supplied sum Table. Over- and under-supplied points during the optimization of the location and the transmitting power Figure 3. Example for an untrained and trained map for the antenna pattern the SOMs only the over-supplied and under-supplied points in the area are considered in the same manner as for the transmitting power. The input vectors ~ X j = ( j G j ) are determined from : G j = T H P t G pj P pj () G pj is the actual gain of the transmitter in the direction j where the path with the highest contribution to the overall fieldstrength of point j is launched. P t is the actual transmitting power of the respective transmitter and P pj the actual power at point j. For the BMU only the angle of the input vector ~ X is considered and only the component G of the weight vector ~w is updated. Cycle under-supplied over-supplied sum Table 2. Over-and under-supplied points during the optimization of the antenna patterns The urban area to be planned is x meters and the resolution is 5m. In tables and 2 the under- and oversupplied number of points after each step is shown. A constraint is that the transmitters are placed below the roof tops a situation characteristic with micro cells. With the equations in section 4. the number of cells is derived as 6, so the SOM for the optimization of the location and transmitting power has 6 neurons. One can see in table and 2 that by optimizing the transmitter locations, transmitting power and

5 the antenna patterns the number of over- and under-supplied points decreases. Exept for the first step and the penultimate step the sum of points that are under- or over-supplied decrease continuously. The configuration thus achieved might be a local optimum. In this case 2 optimization steps have been found to be sufficient. The 2 optimization steps take about 5 minutes on a computer operating at 7 MHz with a 256 MB RAM. In the appendix the isolated cells and the corresponding antenna pattern are shown. [3] R. Menolascino and M. Pizarroso. Software Tools for the Optimisation of Resources in Mobile Systems. In STORMS, Apr [4] J. Vasanto, J. Himberg, E. Alhoniemi, and J. Parhankangas. Self-organizing Map in Matlab: the SOM Toolbox. In Feb. 2. [5] A. Zell. Simulation Neuronaler Netze. Addison-Wesley, 994. A. Appendix Figure 4. Optimization with 6 cells Sum > 3. > 2. >. >. > 9. > 8. > 7. > 6. > 5. > 4. > 3. > 2. >. _< > 3. > 2. >. >. > 9. > 8. > 7. > 6. > 5. > 4. > 3. > 2. >. _<. 6 Conclusions Figure 5. Coverage Cell The results given above indicate that optimizing the location, the transmitting power and the antenna patterns with the kind of neural networks described, is highly suitable for radio network planning. Due to the fast and accurate propagation model, which is also based on a neural network, the optimization process requires only a few minutes. Additional parameters for the planning like the delay spread and interference, which also plays an important role in CDMA systems, should be included in the process in future work. Since the results of the optimization processes might represent local optimums, all methods appropriate to enhance the probability of reaching the global optimum should be exploited. References ,9,8,7,6,5,4,3,2, 8 Figure 6. Antenna Pattern Cell [] T. Binzer, G. Wölfle, R. Hoppe, and F. M. Landstorfer. Dominant Ray Paths for the Planning of Urban Radio Networks. In 9th COST 259 MCM-Meeting in Leidschendam/Den Haag, Netherlands, COST 259, TD (99) 4, Sept [2] T. Binzer, G. Wölfle, R. Hoppe, and F. M. Landstorfer. Sophisticated Neural Networks for the Planning of Urban Radio Networks. In th COST 259 MCM-Meeting in Bergen, Norway, COST 259, TD () 5, Apr. 2.

6 Figure 7. Coverage Cell 2 > 3. > 2. >. >. > 9. > 8. > 7. > 6. > 5. > 4. > 3. > 2. >. _< ,9,8,7,6,5,4,3,2, 8 Figure. Antenna Pattern Cell ,9,8,7,6,5,4,3,2, > 3. > 2. >. >. > 9. > 8. > 7. > 6. > 5. > 4. > 3. > 2. >. _<. Figure 8. Antenna Pattern Cell 2 Figure. Coverage Cell > 3. > 2. >. >. > 9. > 8. > 7. > 6. > 5. > 4. > 3. > 2. >. _< ,9,8,7,6,5,4,3,2, Figure 9. Coverage Cell 3 Figure 2. Antenna Pattern Cell 4

7 > 3. > 2. >. >. > 9. > 8. > 7. > 6. > 5. > 4. > 3. > 2. >. _<. Figure 3. Coverage Cell ,9,8,7,6,5,4,3,2, ,9,8,7,6,5,4,3,2, Figure 4. Antenna Pattern Cell 5 Figure 6. Antenna Pattern Cell > 3. > 2. >. >. > 9. > 8. > 7. > 6. > 5. > 4. > 3. > 2. >. _<. Figure 5. Coverage Cell 6