Surrogate Neural Model for Joint Delay, Hop Count, and Fallback Prediction in LoRaWAN Networks

Low Power Wide Area Networks (LPWANs), such as LoRaWAN, are a foundational technology for large-scale Internet of Things (IoT) deployments due to their long-range communication and low energy consumption. However, multihop extensions of LoRaWAN introduce significant scalability and performance analysis challenges, primarily due to the high computational cost associated with simulating complex network behaviors. To address this, we propose a lightweight Multi-Output Neural Network (MONN) surrogate model capable of concurrently predicting key performance metrics: end-to-end delay, hop count, and fallback path activation. The model architecture utilizes shared hidden layers for feature extraction, followed by task-specific branches to output both regression and classification results. Our model achieves state-of-the-art performance with R<sup xmlns:mml="http://www.w3.org/1998/Math/MathML" xmlns:xlink="http://www.w3.org/1999/xlink">2</sup> scores of 0.9984 for delay prediction, 0.9691 for hop count estimation, and 100% accuracy in binary classification of fallback scenarios. Additionally, the surrogate model offers real-time inference with an average insertion time of just 0.1267 seconds and maintains a compact memory footprint of 226.4 KB, making it highly suitable for deployment on resource-constrained IoT devices. By eliminating the need for full-scale simulations, MONN provides a scalable, efficient, and accurate method for network performance evaluation and optimization in dynamic and large-scale LoRaWAN environments.