Decarbonisation and Machine Learning-based energy monitoring of Natural Gas City Gate Stations

The Italian Natural Gas Distribution Network (NGDN) includes thousands of NG metering and pressure reduction stations, called City Gate Stations (CGS), for injecting gas into low-pressure networks from high-pressure transport networks. These plants are mainly based on the constant-enthalpy throttling of the gas flow to reduce its pressure, which leads to a significant reduction of its temperature by the Joule-Thomson effect. Gas preheating systems that avoid excessive cooling are installed upstream of pressure reduction valves and usually exploit conventional gas boilers. For this process, not negligible amounts of thermal energy are required, usually obtained by burning part of the gas flow rate. In addition to the necessary and rightful urgency of reducing the carbon footprint of the natural gas infrastructure, the objective of containing the NG consumption could also help the system to be more resilient to price variations which in recent months have reached unprecedented levels, hurting the European economy and exacerbating energy poverty condition. This Ph.D. thesis work pursued two different approaches to address this challenge and to contribute to its resolution: the development of white-box models, i.e., based on the first principles of thermodynamics, to carry out techno-economic feasibility analyses and propose solutions to decarbonize preheating demand by lowering natural gas consumption, and, on the other hand, the development of black-box, or data-driven, models to replicate the behavior of the analyzed systems and thus build digital twins that can be useful for monitoring the energy consumption of the systems themselves, as well as for evaluating the impact of possible energy efficiency solutions. An ad hoc thermodynamic model was developed to estimate the thermal energy demand for preheating, exploiting experimental data from various real plants, and simplified models of heat pumps and renewable systems and economics indexes were exploited for techno-economic assessments. Several regression-based machine learning models were developed and trained, a Multiple Linear Regression (MLR) and Artificial Neural Network (ANN) algorithms. The selection of the inputs for these models was obtained through the features selection and engineering process. The algorithms were trained on a training data set relating to a complete year of operation for two real plants and their predictive performances were tested on another data set relating to the subsequent operating conditions of the same plants. The algorithms performed greatly, in terms of all the metrics chosen for their evaluation on the testing dataset. The models were integrated into an energy monitoring system using the CUSUM technique and it was possible through these to identify malfunctions and waste within the testing period. This method allowed for the creation of "baseline" consumption models and to evaluation of the performance of the CGS over time using the CUSUM technique to find the variations between the actual and the modeled gas consumption, being an essential tool for monitoring the effectiveness of the natural gas preheating system. With this thesis work, methods and tools have been developed for the techno-economic analysis of efficiency and decarbonization solutions and machine learning-based monitoring of CGS for NG reduction. This will allow support and help DSOs, providing them with a series of tools that can be useful to address, acting immediately, the challenge of decarbonizing gas infrastructures.