[Paper Review] Cellular Traffic Prediction with Recurrent Neural Network

Shan Jaffry|arXiv (Cornell University)|Mar 5, 2020

Traffic Prediction and Management Techniques18 references4 citations

TL;DR

This paper proposes a Long Short-Term Memory (LSTM) recurrent neural network for cellular traffic prediction, comparing it against ARIMA and feedforward neural networks (FFNN). Results show LSTM achieves higher prediction accuracy with significantly faster convergence, especially on small datasets, making it ideal for real-time, data-efficient traffic forecasting in future 6G networks.

ABSTRACT

Autonomous prediction of traffic demand will be a key function in future cellular networks. In the past, researchers have used statistical methods such as Autoregressive integrated moving average (ARIMA) to provide traffic predictions. However, ARIMA based predictions fail to give an exact and accurate forecast for dynamic input quantities such as cellular traffic. More recently, researchers have started to explore deep learning techniques, such as, recurrent neural networks (RNN) and long-short-term-memory (LSTM) to autonomously predict future cellular traffic. In this research, we have designed a LSTM based cellular traffic prediction model. We have compared the LSTM based prediction with the base line ARIMA model and vanilla feed-forward neural network (FFNN). The results show that LSTM and FFNN accurately predicted future cellular traffic. However, it was found that LSTM train the prediction model in much shorter time as compared to FFNN. Hence, we conclude that LSTM models can be effectively even used with small amount of training data which will allow to timely predict future cellular traffic.

Motivation & Objective

• To address the limitations of traditional statistical models like ARIMA in predicting highly dynamic cellular traffic patterns.
• To evaluate the performance of deep learning models—specifically LSTM and FFNN—against ARIMA for cellular traffic forecasting.
• To assess the training efficiency and prediction accuracy of LSTM and FFNN under varying training data sizes.
• To demonstrate that LSTM can achieve high-accuracy predictions with minimal training data, enabling timely network resource optimization.

Proposed method

• The study implements an LSTM-based model with two hidden layers: the first with 50 LSTM cells and the second a dense layer with a single unit.
• The model uses the standard LSTM cell architecture with forget, input, and output gates, employing sigmoid and tanh activation functions.
• Core equations include the forget gate $\Gamma_f^t = \sigma(W_f[a^{t-1}, X^t] + b_f)$, input gate $\Gamma_u^t = \varphi(W_u[a^{t-1}, X^t] + b_u)$, and output gate $\Gamma_o^t = \sigma(W_o[a^{t-1}, X^t] + b_o)$.
• The cell state is updated via $C^t = \Gamma_f^t \ast C^{t-1} + \Gamma_u^t \ast \varphi(C_u^t)$, and the hidden state is computed as $a^t = \Gamma_o^t \ast \varphi(C^t)$.
• The final output is generated using $y^t = \sigma(W_y a^t + b_y)$, with mean absolute error used to compute training and validation losses.
• All models were trained and evaluated using real-world Call Detail Record (CDR) data, with training conducted in Python using Keras and TensorFlow, and data preprocessing done in MATLAB.

Experimental results

Research questions

• RQ1 Can LSTM-based models outperform traditional ARIMA models in predicting cellular traffic patterns?
• RQ2 How does the prediction accuracy of LSTM compare to that of a vanilla feedforward neural network (FFNN) under the same training conditions?
• RQ3 How does model performance vary when training data size is reduced, particularly in terms of convergence speed and prediction accuracy?
• RQ4 Can LSTM achieve high-accuracy predictions with minimal training data, making it suitable for real-time network planning?

Key findings

• LSTM and FFNN both achieved high prediction accuracy, closely matching ground truth data across all training set sizes.
• LSTM converged to near-zero training and validation error in just 2 epochs when using a medium-sized training set (7142 samples), while FFNN required over 10 epochs.
• With a reduced training set of 3571 samples, LSTM maintained accurate predictions and converged to zero error in under 5 epochs, whereas FFNN required more than 10 epochs and showed higher error.
• When training data was further reduced to 892 samples, the FFNN's performance degraded significantly, with error remaining high even after 20 epochs, while LSTM maintained accurate predictions with error converging to zero.
• The ARIMA model provided a reasonable baseline but failed to exactly match the complex, dynamic traffic patterns observed in the ground truth data.
• LSTM demonstrated superior training efficiency and robustness to small training data, making it well-suited for real-time, data-constrained cellular network forecasting.