Forecasting Demand and Supply Curves in Day-Ahead Electricity Markets: A Machine Learning Approach

The electricity day-ahead market (DAM) plays an important role in ensuring efficient electrical energy distribution and price setting across mostly deregulated power systems. A key feature of this market is the set of supply and demand curves that are constructed from aggregated bid and offer data. These curves are challenging to forecast due to the heterogenity in their construction across days and hours, their inherently monotonic structure. Additionally in quite a few markets, there could be a constraint for forecasting on multi-day horizons. This thesis contributes a compilation of forecasting models and techniques designed to address these challenges, with a focus on robust and interpretable forecasting of these curves. The first contribution is the introduction of a modular framework called the BME-model, which tokenizes market curves into three structurally significant points, B, M, and E, and identifies the significant segment of the curve, for forecasting. We propose two variants, a "pure" variant and a hybrid "combined" variant. These models reconstruct the full curve from individually forecasted tokens and offer improvements over the benchmark naive models. Building upon this, a non-linear extension of the BME-model, using a monotonic autoencoder, is proposed. This approach replaces linear interpolation with step interpolation and uses the encoding of the significant segment of the curve to capture the non-linear structural patterns. An Enhanced Echo State Network (EESN) is also proposed, tailored specifically to forecast both BME tokens and autoencoder encodings with better performance than the previous linear BME-model. To assess model performance in this heterogeneous curve setting, a metric, Heterogeneous Curve Mean Absolute Error (HC-MAE), is proposed. This loss function allows consistent and fair comparison of curves with heterogenous structures. Additionally, an I-Spline-based model is also developed to functionally parameterize the curves while preserving monotonicity. A combination framework is also proposed to combine curve forecasts from different curve forecasting models. The I-Spline-based model and the combination framework were used for a seven day ahead forecasting. Overall, this thesis shows that explicitly modeling and accounting for curve heterogeneity, enforcing monotonicity, and using combinations various models significantly improves the accuracy and robustness of day-ahead market curve forecasts in one-day ahead and seven-days ahead forecasting. These contributions can contribute to the literature for more reliable electricity DAM forecasting.