Neural networks for state-of-the-art modeling tools.
In May 2017 we rolled out a groundbreaking project. For the first time in the gas industry and in Europe we wanted to leverage big data and advanced analytics to model and forecast gas delivery. 5 months later, a neural network based predictive tool was operational.
A few indicators on what we achieved
Balancing the system is crucial for the market: not just physically but also from a commercial point of view. Network operators improvements in information flow towards market players and regulatory bodies support liquidity and transparency. We applied state-of-the-art machine learning techniques to deliver the most reliable and transparent information to all users.
Perfomance improvement
-24% Average error
-33% Maximum error
-45% Errors > 10%
Machine Learning
Machine Learning algorithms
The model is trained to forecast target variables and evaluate the impact of each explanatory variable
It performs a “feature selection” on the explanatory variables reducing their number
The objective function is determined by an algorithm only on the basis of the data structure
Relations between variables are determined by non deterministic models
INPUT DATA
WEATHER
Multidimensional segmentation of the country according to temperature, altitude, humidity, precipitations, etc.
IN-OUT POINTS
Daily time series of provisional and definitive data
CALENDAR
Including the ability to capture critical non standard dates such as bank and floating holidays and potential long weekends
SCADA DATA
Hourly time series of SCADA data points
Delivery forecast at D-2
(updated hourly)
d-2
d-1
d
Forecast
Total delivery
+/- 5% Total delivery