Backtesting coordinated hydropower bidding using neural network forecasting

Abstract

A stochastic programming model for a price-taking, profit-maximizing hydropower producer participating in the Nordic day-ahead and balancing market is developed and evaluated by backtesting over 200 historical days. We find that the producer may gain 0.07% by coordinating its trades in the day-ahead and balancing market, compared to considering the two markets sequentially. It is thus questionable whether a coordinated bidding strategy is worthwhile. However, the gain from coordinating trades is dependent on the quality of the forecasts for the balancing market. The limited gain of 0.07% comes from using an artificial neural network prediction model that is trained on historical data on seasonal effects, day-ahead market price, wind and temperature forecasts. To quantify the effect of the forecasting model on the gain of coordination, we therefore develop a benchmarking framework for two additional prediction models: a naive forecast predicting zero imbalance in expectation, and a perfect information forecast. Using the naive method, we estimate the lower bound of coordination to be 0.0% which coincides with theory. When having perfect information, we find that the upper bound for the gain is 3.8% which indicates that a substantial gain in profits can be obtained by coordinated bidding if accurate prediction methods could be developed.