Rolling Horizon Simulator for Evaluation of Bidding Strategies for Reservoir Hydro

Abstract

In this paper, a practical approach for benchmarking different bidding strategies towards the day-ahead market has been evaluated. A rolling horizon simulation framework is developed and closely integrated in the daily operations of a hydropower producer. The power producer's existing framework of decision support models and data for prices and inflow has been used to simulate the use of alternative strategies on a real life case. In the simulation procedure, a mixed-integer stochastic optimization model is used to determine the bids to the electricity market and the production schedule. It has been demonstrated that simulation over a long timehorizon can be used to evaluate different bidding strategies. Results from the case study show that one single strategy not necessarily will be the optimal one under all conditions, because the optimal strategy will depend on the the state of the system.

A gradient boosting approach for optimal selection of bidding strategies in reservoir hydro

“…In [29], a method for measuring performance of individual historical bidding days has been proposed, where Π , in Eq.1 represents a measure of value for bidding strategy s on day d, and Π , describe the optimal value for the relevant bidding date based on a deterministic strategy with perfect foresight of price and inflow.…”

“…It is also possible to have a skewed threshold where we are more risk averse in relation to choosing one strategy opposed to another. If one strategy is to be selected, previous analysis [29] have shown that a stochastic strategy should be chosen. If a threshold of for instance 0.6 is applied, a clear tendency in direction of deterministic bidding is required before this strategy is chosen in favour of a stochastic strategy.…”

“…An important prerequisite when building a learning model, is the access to data, and specifically in this case, historic bidding performance. To benefit from previous work, this analysis is based on the results obtained in [29] where bidding performance for a river system located in South-Western Norway has been analysed for 2018. Given the large variations that are associated with operation of a hydro-based system, one year of data is considered to be too short.…”

“…In the following sections, results from a case study investigating historical performance of two different bidding strategies will be presented [29]. Further, the steps associated with the machine learning process applied in this article will be described together with examples of application on the historical data.…”