💻 computer science

Demand Forecasting for Internet Bandwidth using Machine Learning to Enhance Resource Management in Nepal

This research proposes the development and deployment of advanced machine learning models to accurately forecast internet bandwidth demand in Nepal's Kathmandu Valley, enabling ISPs to optimize resource allocation, mitigate network congestion, and improve overall service quality through data-driven decision-making.

Original authors: Sachin Maharjan, Shankar Ghimire

Published 2026-08-20

📖 4 min read☕ Coffee break read

CC BY 4.0

Original authors: Sachin Maharjan, Shankar Ghimire

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

The internet is the invisible infrastructure that powers modern life, connecting people to information, business, and each other. Behind the scenes, Internet Service Providers act as the gatekeepers of this flow, managing the massive streams of data that travel through their networks. However, these networks have a physical limit to how much information they can carry at any single moment. When too many people try to use the internet at once, the pipes become clogged, leading to slow speeds, dropped connections, and a frustrating experience for everyone. The challenge for providers is to predict exactly when and how much data will be needed so they can prepare their networks in advance. If they guess too low, the network crashes; if they guess too high, they waste money on expensive equipment that sits idle. To solve this, researchers are turning to machine learning, a branch of computer science where software learns from past patterns to make better guesses about the future, rather than relying on simple averages or rigid rules.

In the Kathmandu Valley of Nepal, where internet demand is growing rapidly, a team of researchers set out to find the most reliable way to forecast this bandwidth usage. They focused on a specific question: which method can best predict how much data users will consume on any given day? To answer this, they worked with eighteen months of real-world data from Dish Media Network, a major provider in the region. This data was a detailed log of every connection, recording exactly when users logged on, how long they stayed connected, and how much information they downloaded. The researchers cleaned this massive dataset, filling in gaps where logs were missing and organizing the information by day, month, and hour to see the big picture of how the network was being used.

The team tested three different approaches to see which one could make the most accurate predictions. The first two methods were traditional statistical tools often used for forecasting. One was designed to spot simple trends, while the other added a layer to account for repeating seasonal patterns, such as higher usage during certain times of the year. The third method was a more advanced type of artificial intelligence known as a Long Short-Term Memory model. Unlike the traditional tools, which look for straight lines and regular cycles, this model is built to understand complex, shifting behaviors in data, much like how a human might learn to anticipate a friend's habits by watching them over a long period.

When the researchers compared the results, the differences were clear. The traditional methods provided a baseline, but they struggled to capture the full complexity of how people actually used the internet. The model that accounted for seasons performed slightly worse than the simpler trend model, suggesting that the seasonal patterns in this specific network were not strong enough to improve the forecast. The advanced artificial intelligence model, however, outperformed both. It produced the smallest gap between what it predicted and what actually happened, with an error rate significantly lower than the other two. This indicated that the neural network was better at recognizing the subtle, non-linear ways that user behavior changes over time.

The study also uncovered what actually drives these changes in usage. The number of active users was the single most important factor, showing a strong link to how much data was consumed. The time of day also played a crucial role, with usage spiking between late morning and early afternoon, while the specific day of the week mattered far less. These findings suggest that for providers in this region, focusing on the total number of users and the time of day is more valuable than trying to predict based on the calendar date alone.

Beyond the technical results, the researchers emphasized the importance of using these tools responsibly. They noted that for such systems to be fair and trustworthy, the data used to train them must represent a wide variety of people and regions. If the data is biased, the predictions could lead to unfair service decisions. The authors argue that with careful oversight, clear documentation, and regular checks for bias, these powerful tools can help providers manage their networks efficiently without compromising the quality of service for their customers. The study concludes that while older methods have their place, the more sophisticated machine learning approach offers a significant step forward in keeping the internet running smoothly for everyone.

Technical Summary: Demand Forecasting for Internet Bandwidth using Machine Learning to Enhance Resource Management in Nepal

Problem Statement

Internet Service Providers (ISPs) in Nepal, particularly in the Kathmandu Valley, face significant challenges in accurately forecasting internet bandwidth requirements. Current demand-prediction techniques are frequently imprecise, leading to ineffective resource management, network congestion during peak hours, and degraded service quality. ISPs are often forced to choose between wasteful resource allocation and service-quality degradation. The core problem is the inability of conventional methods to capture complex, non-linear usage patterns and seasonal fluctuations inherent in internet traffic, necessitating advanced machine learning (ML) solutions to enable data-driven decision-making and optimal network optimization.

Methodology

The study employs a desk-research methodology to investigate an agile data strategy, while the empirical bandwidth prediction work utilizes a specific 18-month dataset from Dish Media Network (DMN), a leading ISP in Nepal. The technical workflow involves the following stages:

Data Acquisition: The study utilizes 18 months of Authentication, Authorization, and Accounting (AAA) data provided by Dish Media Network (DMN). The dataset includes RADIUS logs containing session IDs, IP addresses, connection times, and data consumption metrics (acctinputoctets and acctoutputoctets).
Data Preprocessing:
- Cleaning: Critical columns were selected, and missing data points were imputed using the mean of existing GB_in (bandwidth input) entries to ensure continuity.
- Feature Engineering: The raw second-by-second data was aggregated into daily totals. A new time-indexed DataFrame was created by consolidating year, month, and day columns into a single datetime object.
- Exploratory Data Analysis (EDA): The study analyzed trends, seasonality, and hourly/daily consumption patterns. Key findings included a sharp increase in usage starting November 2022, a seasonal peak in December, and daily peaks between 10 a.m. and 4 p.m.
Model Development: Three time-series forecasting models were implemented and compared using Python (libraries: pandas, NumPy, scikit-learn, TensorFlow, statsmodels):
- ARIMA (AutoRegressive Integrated Moving Average): Configured with parameters $(p, d, q) = (1, 0, 1)$ based on Dickey–Fuller tests and autocorrelation plots.
- SARIMA (Seasonal ARIMA): Extended to handle seasonality with parameters $(P, D, Q, S) = (1, 0, 1, 12)$ , assuming a 12-month seasonal cycle.
- LSTM (Long Short-Term Memory): A deep learning approach using a sequential neural network with 50 LSTM units, a ReLU activation function, and an Adam optimizer. The model utilized a sequence length of 5 (using the past 5 days to predict the next) and was trained for 200 epochs.
Evaluation: Model performance was assessed using Root Mean Squared Error (RMSE) on a held-out test dataset.

Key Contributions

Comparative Analysis: The paper provides a direct empirical comparison of traditional statistical models (ARIMA, SARIMA) against a deep learning approach (LSTM) specifically for internet bandwidth forecasting in the context of a developing nation's ISP.
Identification of Optimal Model: The study identifies LSTM as the superior method for this specific dataset, demonstrating its capability to handle the non-linear and complex temporal dependencies of internet traffic better than linear statistical models.
Ethical Framework: Beyond technical metrics, the paper addresses the ethical implications of predictive modeling for ISPs, emphasizing the need for diverse data representation, regular bias auditing, transparency, and fairness to maintain user trust.
Operational Insights: The analysis highlights specific drivers of bandwidth demand in the Kathmandu Valley, identifying the number of active users (correlation of 0.71) and time of day as the most significant factors, while noting that specific days and years have less short-term influence.

Results

The performance of the three models was evaluated based on RMSE:

ARIMA: Achieved an RMSE of 48,171.0554. While it captured linear trends, it struggled with seasonal fluctuations.
SARIMA: Achieved an RMSE of 50,298.0669. Despite being designed to handle seasonality, it performed slightly worse than the standard ARIMA model in this context, suggesting the dataset's seasonal patterns were not strong enough to benefit the model or that the model configuration was suboptimal for this specific data.
LSTM: Achieved the lowest RMSE of 40,977.3259. The LSTM model significantly outperformed both statistical models, demonstrating its superior ability to capture complex, non-linear patterns and long-term dependencies in the bandwidth consumption data.

Significance and Claims

The paper claims that while traditional models like ARIMA and SARIMA are useful for basic, stationary forecasting tasks, they are insufficient for the highly dynamic and complex nature of modern internet bandwidth data. The study concludes that LSTM models offer a significant improvement in prediction accuracy, making them the most effective tool for ISPs seeking to optimize resource allocation and mitigate congestion.

The significance of the work lies in its demonstration that adopting sophisticated machine learning techniques can transform network management from reactive to proactive. By accurately forecasting demand, ISPs can better upgrade networks, manage congestion, and create tailored packages. Furthermore, the paper asserts that the deployment of such models must be accompanied by rigorous ethical considerations regarding data privacy, fairness, and transparency to ensure equitable service delivery and maintain stakeholder trust. The study serves as a foundation for future work involving hybrid models, external factor integration (e.g., weather, holidays), and real-time forecasting adaptations.