pgautam0033/5g-delay-predictor

language: python tags: 

• 5g
• networking
• cellular-latency
• time-series-forecasting
• onnx
• lstm metrics:
• mae
• r2 pipeline_tag: time-series-forecasting

5G Delay Timeseries Prediction (Hybrid LSTM)

This repository contains a high-performance Hybrid LSTM sequence model optimized and exported into ONNX format. It is purpose-built to predict packet-level end-to-end IP layer delay over 5G cellular network infrastructures while strictly mitigating multi-step error accumulation. 

Track the live prediction dashboard here: https://forecasting-5g.vercel.app 

Project Overview & Research Findings

This project focuses on predicting delay times in 5G networks using time series analysis and machine learning techniques. Our goal is to understand and forecast network performance, which is critical for enhancing the user experience in latency-sensitive 5G services (such as AR/VR, industrial automation, and edge routing). 

Key Insights

By studying packet traces across the IP, RLC, MAC, and Physical layers, the project identifies the primary cause of the triangular pattern observed in Delay vs. SN (Sequence Number) graphs: 

• The Core Finding: The variation is mathematically driven by Frame Alignment Delay at the RLC layer
• (𝑟𝑙𝑐.𝑖𝑛_𝑡−𝑖𝑝.𝑖𝑛_𝑡).
• The Phenomenon: Packets arriving at the IP layer must wait for the next scheduled 5G transmission slot/frame boundary. This waiting time compounds continuously across a frame and drops back down to near zero instantly when a new transmission frame begins, creating a distinctive saw-tooth profile.
• Traditional methods fail to capture these sharp discontinuities, whereas sequence-aware neural networks significantly improve prediction accuracy.

Model Architecture & Hybrid Approach

The time series forecasting model is combined with machine learning techniques to leverage the benefits of both worlds: 

1. Time Series (LSTM): Captures temporal dependencies, trends, and seasonality in sequence arrays.
2. Machine Learning Strategy: Prevents recursive feedback drift by chunking evaluation streams.

Core Component Layer

ComponentDetails

---

Model TypeLSTM (Long Short-Term Memory)

---

LSTM Units50 Units (Stacked 100 → 50 in Deeper Variant)

---

Output LayerDense (1 unit)

---

OptimizerAdam

---

Loss FunctionMean Squared Error (MSE) / Huber Loss

---

Sequence Length30 timesteps lookback

---

NormalizationMin-Max Scaling

Prediction Strategy Comparison

• Pure Recursive: Error accumulates continuously across steps, leading to visible phase shifts, lag, and large cumulative tracking drift.
• Hybrid Approach (This Work): Predicts in small future blocks of 10 packets, then stops and resets the lookup window with real ground-truth data measurements regularly. This bounds error amplification and drastically minimizes phase drift.

Dataset Information

• Source: ExPECA Testbed, KTH Royal Institute of Technology, Sweden
• Size: 40,000+ empirical packet records of network delay times
• Features: Sequence numbers (sn), IP layer timestamps (ip.in_t/ip.out_t), RLC layer timing, and nested MAC/PHY packet trace metrics
• Data Splits: TR1.csv (training) and TE1.csv (testing)

Model Performance & Production Results

The model achieved highly accurate tracking performance on the testing dataset (TE1.csv). It cuts residual variance by roughly 45% over fully recursive baselines and matches tight network service level thresholds: 

Definitive Test Metrics

MetricProduction Evaluation Value

---

Accuracy Score****94.68% (0.9468)

---

**

R2cap R squared 𝑅2 Coefficient**0.9067

---

Mean Absolute Error (MAE) 0.20087 ms (Normalized Scale) / ~0.14 ms to 0.16 ms

---

Root Mean Square Error (RMSE) 0.441362 (Normalized Scale)

---

Precision (At 20.0 ms Threshold) 1.0000

---

Recall (At 20.0 ms Threshold) 0.77657

Note: Deployed in production on Render, a complete streaming batch calculation for 3,470 packet packets delivers live inference within 2.5 seconds while utilizing an optimized ONNX Runtime configuration that dropped background RAM footprint from 450MB down to 80MB. 

Project Integrations

• Live UI Dashboard Site: https://forecasting-5g.vercel.app
• Live Production API Gateway: Email on pgautamlinkedin@gmail.com