Background
Natural gas is the most important energy source in Italy, fueling domestic heating, industrial facilities, and thermoelectric power plants. The transport of natural gas is guaranteed by a network of over 32 thousand kilometers of pipelines, with the activity entrusted to the transmission system operator (TSO). According to government regulations, the TSO must also provide gas demand forecasts to ensure safe, reliable and efficient operational planning and drive gas prices as well as future investment requirements. This task results particularly difficult in Italy due to the complicated structure of the gas network as well as volatile weather patterns.
The TSO initially generated forecasts using time series models. However, due to the increasing complexity of the gas network infrastructure and protocols, it became apparent that a new and more advanced approach was required.
Decisions and Actions
Our proposal was to develop a forecasting system capable of increasing the accuracy and stability of Italian gas demand forecasts through a Machine Learning approach. The main focus was on the one-day-ahead model, subject to an economic incentive proportional to the daily percentage error.
The system:
provides hourly predictions of the volume of gas delivered across Italy for the following 4 days, with detail per type of usage (Thermoelectric, Industrial and Civilian);
utilizes a set of neural network models that consider real-time gas usage, weather forecasts, historical Company data and electricity demand forecasts;
has a rolling operating mode, with the models continuously updated with the most recent data.
Results
The forecasting system has brought the following benefits to the client:
increased efficiency of gas transportation activities, due to the control room taking better actions to physically balance the network;
improvement of the quality of the information provided to the market about the imbalance of the system, leading to more accurate market interventions;
performance boost of over 40% compared to the previous models, with the day-ahead model achieving a 2.9% yearly average percentage error and significant economic subsidies.