116. An approximate dynamic programming approach to the control of regenerative power systems
Invited abstract in session TC-9: Operational planning in energy systems, stream Energy and Sustainability.
Thursday, 11:45-13:15Room: H15
Authors (first author is the speaker)
| 1. | Florian Scholz
|
| Institute of Management and Economics, Clausthal University of Technology | |
| 2. | Christoph Schwindt
|
| Institute of Management and Economics, Clausthal University of Technology |
Abstract
In this talk we consider a power supplier operating regenerative power plants and electricity storages. To meet the customer demands, the supplier can feed electricity generated by the power plants, released from the storages, or procured on a continuous intraday market. The supplier can also use the storages to achieve trading profits. The respective decisions must be taken immediately before the load, the utilization level of the power plants, and the market price are known. For each of these three uncertain parameters, the supplier has forecasts with known accuracies available. If the load is not entirely satisfied, the supplier bears the cost for the balancing services. The resulting control problem is modeled as a discrete-time Markov Decision Problem (MDP) on a time horizon comprising the 96 quarter hours of one day, where the states coincide with the storage levels and the actions represent procurement and sales volumes.
We propose a heuristic solution method starting from an LP formulation of the average-value certainty equivalent. The LP solution is interpreted as a state-independent policy, which serves as an initial solution to the MDP. This solution is then improved using a restricted version of the MDP where only actions within a given range around the initial policy are considered. A discretized version of this problem is efficiently solved by recursively evaluating the standard Bellman equations. We present the results of a computational experiment where based on past prices and forecasts for loads and wind speeds we generate price forecasts using a neural network trained on historical data. By varying technical, environmental, and cost parameters of the problem instances, we analyze their impacts on different performance indicators of the energy system.
Keywords
- Energy Policy and Planning
- Stochastic Models
Status: accepted
Back to the list of papers