FUTURE OF POWER CONSUMPTION
An excerpt from the Master's Degree Thesis in Quantitative Finance at Politecnico di Milano
by Giuseppe Messuti
The use of electricity is now a necessity in everyday life, moving closer and closer to essential commodities. It doesn't only feed mobiles and laptops, but is able to replace domestic gas in the kitchen, oil and diesel for public and private transports, and allows us to complete day-to-day actions providing light every day of the year.
​
Compared to gas, electricity is a passing good, which can be provided in large quantities after request, but it can be hardly stored. Most of the power we are using now has been generated during last 30 minutes, and therefore plants which feed power to the grid should have as soon as possible an range of next power consumption .
This is because activation and shutdown of generation plants request much more time compared to a common switch, and over-generation not only damages their profits, but the grid itself with an power overload. On the other hand, an under-generation leads to a black-out, followed by the launch of emergency power sources, which usually aren't renewable, as thermal, which need shorter activation times.
​
Therefore, a good estimation of future power consumption would allow companies to save money, while cleaning environment from pollution, decreasing CO2 in favour of renewables. Moreover, a decrease of consumption variability around its expectation may encourage creation and development of wind and photovoltaic on-shore and off-shore farms, which should lead to a quick decrease in use of non-renewables for most of power production.
​
In this framework arises the necessity of mathematical models able to carefully and reliably describe electricity consumption.
In this work we focus on the daily electricity consumption process in a region of UK.
We aim to decompose the consumption in a main seasonal shape which takes into account multiple deterministic factors, and a relation with daily weather conditions.
​
Our main objective is the modelling of daily distribution of consumption process, which has variance clustering during winter period, as well as a main seasonal behaviour combined to a long-term trend and autoregressive effects.
In this thesis we model the de-seasonalized process with respect to weather conditions based on linear and non-linear relations, taking into account the following modeling descriptions:
-
a Linear Model Regression;
-
weather conditions distance-based Nearest Neighbour Regression;
-
weather conditions covariance-based conditional distribution with Gaussian Process Regression;
-
high non-linear relations with Auto-Regressive and Feed-Forward Artificial Neural Networks.
We desire to show how the increased complexity in the model description allows to achieve a more accurate consumption forecast.
​
This hybrid approach can lead to meticulous description of the power consumption process, providing in particular:
-
a term-independent forecasting model for expected daily power consumption, which can be used for point-wise predictions;
-
a term-independent simulation environment, that gives a wider description of consumption process.
It is possible to graphically and numerically compare results of described models, experiencing strengths and weaknesses of each of them.
