Information Theoretic Learning Applied to Daily Streamflow Forecast and Its Impact on the Brazilian Hourly Energy Spot Prices.

The Brazilian energy spot price is obtained through a chain of dynamic stochastic optimization models that works with the uncertainties related to its continental hydrogeneration-based power system. In that sense, this paper presents an information theoretic learning neural forecasting model for daily streamflow prediction of Brazilian hydroelectric power plants. More precisely, the maximum correntropy criterion was used as the error function of a multilayer perceptron. After the prediction stage, the generated outputs were used as one of the inputs of the model chain that is used to compute the hourly energy spot price in Brazil. To the best of the authors' knowledge, it is the first paper that aims to analyze the impact of the streamflow prediction on the Brazilian hourly energy spot price formation. In terms of streamflow forecast, results indicated that the predictions originated from the proposed forecasting model were equivalent to the ones from the official models, especially in the first predicted day. In the spot price graph analysis, the main result pointed that the curves provided from the modeled structure were closer to the values obtained using the actual flows than the official prices, which shows that the proposed work could produce prices more aligned to the real hydrological system conditions. From that, the study's relevance is due to the conclusion that the official process of streamflow forecasting can be improved to generated outputs more consistent with the actual system conditions, to avoid further expenses in the system operation due to potential unscheduled hydrothermal dispatches. [ABSTRACT FROM AUTHOR]