Soutenance de thèse : Léonard TSCHORA

Machine Learning Techniques for Electricity Price Forecasting

Doctorant : Léonard TSCHORA

Laboratoire INSA : LIRIS

Ecole doctorale : ED512 Informatique et Mathématiques de Lyon

Electricity is essential for the energetic transition due to the diversity of greenhouse-gas free means of production and its potential to replace fossil fuels. However, it requires constant balance between generation and consumption, and can't be stored efficiently. Thus, it's necessary to use Price Fixing Algorithm (PFA) for developing competitive markets. Daily, Euphemia, determines the prices for the next day. Unlike other speculative markets, the price is algorithmically computed that renders its forcasts paramount for business applications. Electricity Price Forecasting consists in predicting the 24 hourly prices before their fixation at 12am. The literature highlights two incomplete approaches: expert models aim at replicating the PFA and computing the prices based on estimates of its inputs, but fail to produce accurate forecasts in practice. Data driven methods directly estimate prices using exogenous variables and past prices, but lack transparency. Also, the true relationship between variables and prices is only captured by Euphemia, implicitly limiting the performances of data driven approaches. The first challenge is to produce explainable EPF models using Shap Values, a model- agnostic explanability tool. Then, we represent the European network as a Graph where each country is a node labeled with its prices. We estimate the Graph edges using an optimization problem prior to training. With a Graph Neural Networks, we forecast prices for all markets simultanesously. Lastly, we combine the Euphemia algorithm with in a Neural Network (NN) that forecasts its inputs. To consider the price forecasting error in the NN's training, we compute the gradient of Euphemia's output with respect to its input, by vanishing the derivative of the dual function using a dichotomic search. We hope this thesis will be beneficial for the EPF practitioners and we also believe that our work on mixing optimization problems with machine learning models will benefit the broader Machine Learning community.