Electricity and utilities
Forecasting short-term electricity demand remains a prominent topic to utilities companies, due to the importance of future loads in the operation of utility plants, transmission systems and trading on energy markets. Fuelled by the competitiveness of today's deregulated and highly volatile electricity markets, small improvements in forecasting accuracy can account for large operational profits, far exceeding that of one million GBP in profits as an equivalent to1% improvement in accuracy on the UK market reported in the 1980s.
The Lancaster Centre for Forecasting has substantial expertise in applying both traditional statistical methods, as well as novel statistical contenders such as Triple Exponential Smoothing. Additionally we have extensive expertise with algorithms from artificial intelligence such as neural networks and k-nearest neighbours which are capable of capturing the underlying nonlinearities with temperature (in interaction with irradiance, cloud cover, precipitation, and/or wind chill) and able handle big time series data at the same time.
Artificial neural networks, k-nearest neighbour and other advanced nonlinear methods
Artificial neural networks run in over 50 utilities internationally. In addition to statistical models such as traditional linear regression and more complex triple seasonal exponential smoothing methods, we develop start-of-the-art predictive models based on artificial neural networks and k-near neighbour regression which cope with the inherent nonlinear interdependencies in load forecasting, e.g. with temperature and wind speed, in addition to the unknown dynamics of the data generating process. The result is a set of tailored solutions meeting a wide variety of user needs.
Data segmentation and time series clustering
Advanced time series segmentation and clustering techniques allow the design of multiple local models corresponding to homogeneous subsets of the electric load demand. This provides detailed predictions for each segment or cluster of interest, which can then be recombined to produce an improved aggregate prediction at the highest level. An additional benefit of this approach is that the local models are independent, and allowed to vary stochastically over time providing even greater flexibility.
Electricity demand time series data contain complex (multiple) seasonal patterns. This includes the presence of a within-day seasonal cycle visible from the demand profile of one day to the next, a within-week seasonal cycle which appears when one compares demand on corresponding days of the same week, a within-year seasonal cycle which appears when multiple years of data are available. We help companies deal with these difficult to model features.
Semi-automatic identification of intraday outliers using time series clustering
We apply computational and statistical techniques such as time series clustering to facilitate the identification of features of interest including clusters (e.g. demographic zones, household types), outliers, and trends for intraday electricity load demand in a semi-automatic manner, and determining the semantics of such features.
Through the Masters programme at Lancaster University, we facilitate the development of young researchers in the area of Energy and Utilities forecasting conducting a number of Masters student projects. Examples of these initiatives include:
- Control Assessment Criteria Methodology in Multi-Project Programmes, National Grid Transco
- Measuring Substation Complexity, National Grid Transco
- Analysis of the Trends for the Patterns in 15-minute Water Usage Data, Thames Water
The Forecasting Centre regularly engages in general as well as custom-made training courses for teams specializing in electric load and utilities demand forecasting. Past courses included attendees from NPower, British Gas and Scottish Power.
For any enquiries please contact email@example.com.