Economic Optimization and Control of Distributed Energy Resources – Deep Reinforcement Learning as an Option for Automated Real-Time Operation
(Room Arena 5)
20 Jun 18
11:00 AM
-
12:30 PM
Tracks:
Arena 5: Artificial Intelligence and Blockchain – Their role in our market
The efficient operation of distributed energy resources (DER) can pose some severe challenges especially if DER ought to handle signals from a market or a grid operator: a) To successfully integrate DER into larger energy systems (e.g. smart grids, microgrids) and markets (e.g. intraday and balancing market or peer-to-peer trading) it is crucial to enable the distributed units to quickly react to real-time signals. This increases the requirements for the speed of computation of the algorithms used to determine the units’ optimal dispatch. b) More and more measurement data that contains valuable information is available from distributed units. Algorithms should be empowered to appropriately make use of it. c) Compared to centralized systems the optimal dispatch has to be determined simultaneously for a large number of smaller energy systems, each with its own technical particularities. This further increases the computational burden imposed by the necessary optimization algorithms. In the recent past Deep Reinforcement Learning (DRL) has emerged from the field of artificial intelligence and since then solves complex automation tasks across a continuously broadening range of industries. This conference contribution will show how DRL can help to face the described challenges of efficiently operating DER in real-time environments. We present an algorithm that dispatches numerous complex energy systems while frequently receiving near real-time information e.g. on market prices, weather forecasts and measurement data. It will be shown that DRL is well fitted to extract the relevant information from large data streams and deal with the stochastics of prices, demand and renewable production as well as time-varying quality of data and forecasts. The algorithm is compared to conventional optimization methods. It is shown that conventional algorithms fall short either on the quality of results or their capability to interact quickly enough with a real-time system.