education

Education


Round Corner
Department of Computer and Information Science

Oppgaveforslag

LSTM-based models for forecasting intraday power prices (AI Lab Pitch) [2019/2020] [Norwegian OpenAI Lab + Refinitiv]

[Project in collaboration with Norwegian OpenAI Lab]

[Project in collaboration with Refinitiv. Refinitiv (former Financial&Risk division of Thomson Reuters) is one of the largest data and analysis providers in the world for the commodities and financial sector see here)]

Problem Description
The extreme increase of installed renewable capacity in Europe is shifting the focus from day-ahead trading to intraday trading in virtually all countries in Europe. This is due to the fact that predicting wind and solar power output over horizons > 24 hours is still challenging, with mean absolute errors seldom below 10%. For this reason we are exploring the possibility of forecasting intraday prices and volumes for the coming hours, using the most recent weather signals, the latest traded prices, and the information from grid operators.

Data
The problem can be framed in ML terms as a 3-dimensional problem, where for each today’s contract (hour) we measure the development of the weather signal and the development of
the traded price since yesterday. We believe that LSTM might be an interesting technique to test out, but the candidate is welcome to explore other techniques. The candidate must be interested in working with unstructured and possibly noisy and incomplete data, and must take a proactive approach. For more information about modeling intraday prices, please see HERE

Business challenges with the data set:
Power markets are becoming increasingly driven by weather data, because of the immense amount of renewable generation added to the European stack every year (currently ~100 GW solar and ~ 200 GW wind power installed in Europe). Analyze and understand such data in real time is therefore crucial for a well functioning market, and will be of increasingly importance in the future.
Scientific challenges and methods: Currently most literature models are either regression models or linear optimization models. We believe that shallow and deep ML approaches can help identifying patterns that currently are not visible in the data sets.


Expected outcome
Literature review, understanding of the data/problem, simple prototypes, ideas for future developments.

Co-SupervisorGabriele Martinelli, Senior Analyst, European Power @ Refinitiv

 

Faglærer

Massimiliano Ruocco Massimiliano Ruocco
Adjunct Associate Professor
248 IT-bygget
 
NTNU logo