Predicting electricity supply and demand curves with functional data techniques

The profitability of electricity companies in Europe, especially in Spain, has been declining due to the poor performance of liberalized activities (generation and commercialization). This decline is caused by reduced demand, decreased investment, and asset value loss from the economy’s decarbonization. In this context, precise forecasts of hourly supply and demand curves, as well as hourly prices in the wholesale electricity market, are crucial for optimizing energy buying and selling strategies. This work focuses on the daily Spanish spot market, where energy is traded for the 24 h of the following day. This market is crucial as it accounts for the highest volume of energy traded, contributing the most to the final electricity price (88.7% in 2023, according to the Spanish System Operator). Optimizing strategies in this market can significantly improve participants’ economic outcomes. Despite extensive study, there is still room for improvement. This paper proposes predicting hourly supply and demand curves and the matching price and matching energy for the following day using various functional analysis techniques. It combines functional analysis and machine learning techniques, incorporates seasonal and regular lags due to the strong dependency found between consecutive hours, and includes meteorological information from eight variables across Spanish provinces. Additionally, we do not assume smooth curves, leading to more realistic predictions. Finally, predictions are adjusted with the closest training set curve. The extensive backtesting results highlight the importance of considering all these aspects to reduce prediction errors for curves and hourly prices and energies. • Aggregate curves and matching prices/energies are predicted in a two-step method. • The model combines functional analysis techniques and machine learning. • The model considers temporal dependency and meteorological information as features. • We design the model and post-processing to avoid rather smooth predicted curves. • Exhaustive back testing shows good prediction accuracy of curves and prices/energies.