Predictive analytics can be employed by telecommunications operators to gain valuable insights into the network performance and make data-driven decisions in order to optimize the quality of service for the users while saving the network resources. However, predicting the network key performance indicators (KPIs) is a challenging task in large scale cellular networks with complex spatio-temporal variations. Moreover, to optimize, expand, and modify the strategy of mobile networks, many parameters are changed or features are being implemented every day. One of the major challenges in network maintenance is to track the side-effects of such changes, in order to avoid any anomalies, by calculating the amount of degradation or improvement. Sometimes these changes coincide with other network seasonality behaviour, hence, adding complications to the network. Inspired by the promising performance of recurrent neural network (RNN) in time-series prediction, we developed a KPI prediction model using an RNN algorithm in a mobile network with various technologies (3G, 4G, and 5G) and frequencies. Model training is performed on three years of historic mobile network data to achieve a high accuracy model.
The most important characteristic of RNNs is using the memory to process sequences of inputs. It means, unlike the convolutional neural networks (CNN) architecture, where there is no memory to capture the temporal correlation, RNNs have the capabilities to capture the sequential correlations and information. There are three types of recurrent neural networks, the simple RNN, LSTM, and gated recurrent unit (GRU).
PROPOSED RNN DEEP LEARNING MODEL
Fig. 2 shows the architecture of the proposed model. In the first step, the model detects the type of KPI, and based on this detection, the proper hyper-parameters such as the number of epochs, batch size, learning rate, activation function, dropout rate, number of hidden layers and units, loss function, and optimizer are assigned to this system. Then, the dataset is passed through the RNN and Dropout/Dense layers and the prediction trend appears in the output.
In time-series prediction, with given a univariate time-series
including T observations, the N next observations of the time-series
can be predicted. Time-series in different applications or fields can have a variety of resolutions (e.g. hourly, daily, monthly) with considering the T interval. The horizon (N) refers to how far in the future we can predict this trend and it can be in the range of days/weeks/months. In this step, the data should be separated into two parts of training and test datasets. In this project four years of KPI in a mobile operator has been divided into the training and test datasets by 80/20 percentage.
TRAINING THE MODEL
The training part should consider all types of KPIs, thereby we cannot use a unique system for all sorts of KPIs. The hyperparameters should be tuned based on each KPI. Some of the most important parameters are the active function, optimizer, loss function, size of the hidden layer, batch size, or type of RNN (GRU or LSTM). This part of the system has been designed dynamically. It means that a detector finds the types of KPI thus, the model sets the best configuration and hyperparameters for the specific input KPI. Table 1 shows the network configurations with all possible and effective parameters for some KPIs.
In table I, the Loss Functions are abbreviated as following
RESULT AND PRACTICAL EXPERIMENT IN A LIVE MOBILE NETWORK
The final proposed systems with appropriate parameters which have been presented in table I predict the payload, throughput, drop rate, and established success rate. The following figures show the accuracy by depicting the real and predict trends on daily basis.
In this section, we present the performance of our model in a practical experiment. Several changes should be implemented in different parts of mobile networks every day, such as parameter re-tuning, strategy changes in optimization part, or cutover and expansion of fiber or microwave link in transmission department. The most important phase of these activates would be seeing the positive or negative impact of them, to decide whether they should be kept or reverted back. In this experiment, we activated the roaming between two mobile operators for some specific sites and we wanted to know if the traffic increases or decreases in these sites. In general, we do not know whether revenue was gained or lost with this implementation. By using this proposed RNN method, the KPI trend can be predicted to see the impact. Fig. 7 shows the result of this deep learning system on the payload, it illustrates with this implementation (Blue) and without this activation (Orange), in other words, the orange line shows the prediction trend.
The second presented application is the throughput analysis. An optimization action has been taken, and the result of the prediction system in Fig. 8 shows that the throughput improves after this action. It means that without this action, the throughput may be in the vicinity of 8.3 Mbps, but by means of this optimization, the throughput has been improved to 9 Mbps. Therefore, the results recommend to keep this change.