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 in transportation, heating and industries. However, it requires a constant balance between generation and consumption to maintain intensity in the network, and it can’t be stored efficiently. It is then necessary to use Price Fixing Algorithm (PFA) for developing competitive markets. Daily, the European PFA euphemia determines the prices for the next day in Europe, called the Day-Ahead prices, that maximize the Social Welfare, while maintaining energy balance. Unlike other purely speculative markets, the Day-Ahead prices is algorithmically computed. Forecasting them is thus a unique and challenging task.

This introduces the problem of Electricity Price Forecasting (EPF) at the European scale, that consists in predicting the 24 hourly prices for each market before their fixation at 12am. The literature highlights two approaches: Expert models, that aim at replicating the PFA and computing the prices based on estimates of the inputs of euphemia, and Data- Driven methods that directly estimate prices using exogenous variables and past prices. Both approaches are incomplete: Expert models approaches are theoretically appealing but fail to produce accurate forecasts in practice. Conversely, Data-Driven approaches lack transparency, lowering the forecasts reliability. Also, the true relationship between variables and prices is only captured by euphemia, implicitly limiting the performances of Data-Driven approaches.

This thesis addresses those limitations. The first challenge is to produce accurate and explainable models for a given market. We achieve the former by extending methodologies from the literature, while we use Shap Values, a model-agnostic explainability tool, for the latter. Then, we build a multi-market forecasting model by representing the European network as a graph where each market is a node labeled with its prices. Graph edges are connection lines between markets, and we estimate the cross-market flows using an optimization problem prior to training. 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 dichotomy search.

We believe this thesis will be beneficial for the EPF practitioners and will contribute toward bridging the gap between Expert models and Data-Driven approaches. We also believe that our work on mixing optimization problems with machine learning models will benefit the broader scientific community.