Trading with electricity presents great challenges due to its complexity and stochastic characteristics. Nowadays, when establishing new trading strategies, emerging concepts and approaches from data analysis and artificial intelligence are of great help and importance. In this master thesis our goal is to explore different possibilities for using deep reinforcement learning methods as a basis for new electricity trading strategies. Electricity markets are thus modelled as Markov decision processes using deep reinforcement learning methods. For the establishment of an efficient strategy we implement a reinforcement learning agent, which uses historical trade data to learn the optimal trading strategy.
In this master thesis there is a big emphasis on the theory of reinforcement learning, which defines punishments and rewards as two leading concepts during the learning process. There are different approaches inside the reinforcement learning theory and praxis but we focus on the Q learning method. Because the electricity market has an infinite number of possible states, there is a need to widen and upgrade the Q learning method which is addressed with the use of deep neural networks. The result is a deep Q learning algorithm, which is used to train the agent’s trading strategy.
Effectiveness of developed strategy is measured by total profit and other metrics such as average margin and win ratio. In this master thesis we developed an algorithm for electricity trading that provide a learning convergency that is stable and of sufficient speed. Results of trading with implemented algorithm
indicate significant improvements when compared to random trading.
|
