BME Model: Forecasting Electricity Supply and Demand Curves Using Disentangled Prices and Volumes

This paper introduces a novel modeling setup for forecasting day-ahead electricity market demand and supply curves, seen as bid prices as a function of bid volumes. These curves are different for each different next-day hour. The daily data structures on which these curves are assembled have very peculiar features, which make them very difficult to handle. One of these features is that each curve can be seen as defined on a set of volume values that each day and for each hour change in position and number in a non-deterministic way. Thus, modeling, comparing, and forecasting such curves cannot be made in standard econometric ways. In this paper, it is proposed to carry on the forecasting task by looking at the curves as sequences of bid points, and by tokenizing them into three structurally significant special points and a special segment, without relying on the functional approaches that are so common in related electricity finance literature. The modeling setup is mainly based on linear forecasting methods, with a very modest addition of nonlinearity by means of a simple feedforward neural network. The setup is modular and can be used both in what is called a “pure variant” and a “combined variant”. Three standard and important benchmarks, namely the naive and smarter naive models adapted to curves, and the stochastic-functional autoregressive model already present in the literature, are used for comparison. The forecasting ability of the proposed setup is comparatively tested on data from the NORD zone of the Italian IPEX electricity market. Overall, numerical results demonstrate that the “combined variant” of the setup is the most effective forecaster among all benchmarks, for both demand and supply curves, and across all hours.