Previsão de preços e variáveis hidrológicas no mercado livre de energia brasileiro: uma abordagem baseada em machine learning e métodos Bayesianos

Abstract

Agents trading energy in the free market make decisions by considering, among other factors, the future hydrological conditions of the National Interconnected System (SIN), the impact of these conditions on the predictions of the Settlement Price of Differences (PLD) (spot price of energy), and the behavior of prices negotiated in the market, the last summarized by the Forward Price Curve. In the short-term, PLD prospecting studies rely on streamflow rates predicted by a rainfall-runoff model, which uses as input the precipitation data generated by numerical weather forecasting models. In the medium term, agents adopt varying strategies. Some rely on scenarios derived from historical streamflow timeseries, while others integrate a rainfall-runoff model with historical precipitation data or seasonal precipitation forecasts. Regarding the Forward Price Curve, it is typical to consider only its present state, with limited research dedicated to forecast this particular time series. In light of this context and the ongoing need to refine the forecasting of key variables for commercial decision-making, this research aims to develop methodologies to forecast hydrological variables (streamflow and rainfall) and the price established by agents in the Brazilian energy free market (Forward Price Curve). Additionally, this research aims to strengthen the relationship between the academic and business sectors by proposing practical solutions (in the form of products) to real problems of the energy market. The first product focuses on the short-term and introduces a new framework based on Bayesian Model Averaging (BMA) to forecast natural inflows to reservoirs and hydropower plants 7 ______________________________________________________________________________ ______________________________________________________________________________ in the Brazilian hydropower system. BMA is a statistical method that combines models using posterior probability distributions. Streamflow forecasts for 11 days ahead are produced by assigning weights to the members of a Multi-model Ensemble (MME) of rainfall-runoff hydrological models. The weights are dynamic, periodically recalculated with recent data. Selecting high-performing members, using Occam's Razor principle, improves the MME's performance. Tested on 139 reservoirs between 2019 and 2020, the method showed superior forecasts compared to individual models, especially in the South, Midwest, and Southeast regions in the early days of the forecast horizon. Focused on the medium-term case, the second product of this research aims to refine seasonal precipitation forecasts for Brazil by developing predictive hybrid precipitation models using Multivariable Linear Regression (MLR) and Support Vector Machine (SVM). The approach incorporates climate indices related to different teleconnection patterns that affect Brazil precipitation, the Climate Prediction Center (CPC) unified gauge-based global daily precipitation analysis, and precipitation forecasts from Seasonal Forecast System 5 (SEAS5) as predictors. Validated from January 2017 to December 2020, the MLR and SVM models showed greater accuracy and lower bias than SEAS5, especially in the Southeast, Midwest, and North regions during the DJF quarter, with the SVM model demonstrating the best overall performance. The third product introduces a methodology centered around Artificial Neural Networks (ANNs) to forecast the Forward Price Curve of conventional energy in the Brazilian energy free market. This study evaluates Long Short-Term Memory (LSTM), Gated Recurrent Unit (GRU), and Recurrent (RNN) neural networks, both uni- and bidirectional, alongside the traditional Multi-Layer Perceptron (MLP). It also considers structural variations in ANN architectures and the impact of exogenous predictors related to the SIN and the Brazilian energy free market, including natural inflow energy, load, reservoir storage levels, and the energy spot price. Among the neural networks evaluated, the GRU demonstrated superior predictive performance, while the LSTM excelled with regard to accumulated financial return in an experiment where a trading agent performed directional trading informed by neural network predictions.