Pipeline Cathodic Protection
Predictive modeling for proactive corrosion management across Europe's natural gas network
Reactive → Preemptive
System Shift
4,000 km
Pipeline Network
Dual Threshold
Control Challenge
The Problem
Fluxys operates over 4,000 km of natural gas pipeline spanning Belgium and neighboring European countries. Their Impressed Current Cathodic Protection (ICCP) system was entirely reactive. Personnel had to physically travel to measurement points across the network, take readings, and then decide whether to adjust the current being sent through rectifiers. This demanded significant manpower across a geographically distributed network of rectifiers and measurement points spanning the country. The core challenge was a dual threshold: pipeline potential that drifts too high leads to corrosion, while applying too much rectifier current risks hydrogen cracking. Keeping voltage within this narrow safe band reactively was unsustainable at scale.
Approach
Working at OpenHub's 'Garage for Innovators' with an international, multidisciplinary team in Belgium, I built predictive models to enable proactive voltage management. The data was messy. On-potential and off-potential measurements came from the Antwerp area and many other locations, each with its own rectifiers and monitoring points, and this geographic variety made the dataset complex. I gathered external data including weather conditions via web scraping and data providers, along with groundwater information that could affect readings. To find useful predictive relationships, I applied Granger causality testing between different time series such as protection measurements and weather data. Statistical hypothesis testing for stationarity and STL decomposition were essential for handling the time series properties of the data. After exploring traditional ML and various time series approaches, I built stacked LSTMs in TensorFlow/Keras, training separate models for max, average, and min voltage predictions. Each model took triple input features (voltage, current, and change in current) and used a custom spike-weighted Huber loss to penalize missed voltage excursions. For multi-step forecasting, I used a recursive approach where the model predicted one step ahead given planned future current values, appended that prediction to the input sequence, and repeated out to the forecast horizon. Hyperparameter tuning was done with Bayesian optimization via Scikit-Optimize.
Results
The work helped lay the groundwork for Fluxys to build a preemptive cathodic protection system, moving away from their entirely reactive approach. By demonstrating that predictive models could anticipate voltage trends across the network, the project showed a viable path to proactive management that could significantly reduce the manpower required for constant manual measurement campaigns while maintaining pipeline integrity within safe thresholds.