Jakob Tjurlik
In collaboration with 1Optic and Utility Connect, I explored how advanced deep learning models can predict signal anomalies in telecom networks used for smart energy management. While challenges remain, the project provided valuable insights and outlined next steps for improving anomaly prediction, including incorporating environmental data for enhanced accuracy.
The problem: Signal anomalies in smart energy networks
Smart meters, vital for energy monitoring, rely on radio access networks (RANs) for communication. These networks occasionally experience signal anomalies, characterized by prolonged hikes in signal strength at the receiving antennas. Such increases in the received signal can be caused by a variety of factors, including equipment malfunctions, software misconfigurations, or external environmental phenomena.
One such environmental phenomenon, identified by Utility Connect, is tropospheric ducting, where atmospheric conditions like temperature inversions allow radio waves to propagate over large distances. These anomalies disrupt grid reliability, delaying critical data transmission and complicating energy management for utility providers. Recognizing the operational impact, Utility Connect and 1Optic entrusted me with this project to explore how predictive models could help mitigate these signal anomalies.
The methodology: Leveraging neural networks
To tackle the issue, I developed and tested advanced neural network models that analyzed historical signal measurements from base stations across the Netherlands. Various setups were tested to evaluate the predictive power of models such as Convolutional Neural Networks (CNNs), Long Short-Term Memory (LSTM) networks, and hybrid architectures like CNN-LSTM and ConvLSTM.
The goal was to predict signal anomalies up to 12 hours in advance, enabling proactive responses to interference. By focusing on internal network Key Performance Indicators (KPIs) and spatial data in the initial stages, the project aimed to establish a baseline understanding of signal behavior, before considering external variables like environmental data.
Visualization of the Received Signal Strength Indicator measurements (RSSI) over time on a sample base station. A base station typically has three antennas, each with two carriers, ensuring continuous data reception from multiple directions. Signal anomalies appear as large upward hikes in signal strength.
The results: Challenges and opportunities
While the neural networks performed well in general signal forecasting, predicting anomalies specifically remained a significant challenge. Below are some key findings:
- Baseline models hold their ground: A naive persistence model, which repeats recent RSSI values, consistently matched or outperformed deep learning methods in anomaly prediction.
- Complex models, marginal gains: Advanced architectures like CNN-LSTM and ConvLSTM offered limited improvements over simpler models of CNNs or LSTMs.
- Variable selection matters: Incorporating network KPIs and spatial data improved short-term forecasts but didn’t significantly enhance long-term predictions.
These results reflect the complexities of anomaly prediction in dynamic, interference-prone environments. The absence of environmental variables, which became apparent as the project progressed, was identified as a key area for further exploration.
The figure shows a difference between the actual signal indicator and the one predicted four hours ahead by an LSTM network, on a sample antenna and a sample time frame. One time step corresponds to 30 minutes. It shows that the prediction algorithm was not able to forecast the onset of anomalies at a 4-hour horizon, as well as underestimating the magnitude of some higher increases.
The impact and future directions
This project emphasized the importance of collaboration and interdisciplinary insights. Working with 1Optic and Utility Connect provided not only the resources but also the domain expertise to explore this complex problem effectively.
Looking ahead, incorporating meteorological data—such as temperature, atmospheric pressure, and other weather conditions—is a promising next step. By linking environmental factors to signal anomalies, these variables could improve predictions and enable more proactive responses. Together, these advancements will strengthen long-term forecasting of anomalies and ensure more resilient telecom systems for energy management.
About the author
Jakob Tjurlik is a recent graduate at Tilburg University in M.Sc. Data Science & Society. This project was conducted as part of a Master’s thesis in collaboration with 1Optic and Utility Connect.