Bidding Strategic of Virtual Power Plant Based on End-to-End Deep Reinforcement Learning

Apply deep reinforcement learning for bidding strategy generation in electricity market, enables the market participant to obtain accurate feedback with the bidding action on the impact of market clearing outcomes, then optimize the action to be more profitable according to the feedback. However, current deep reinforcement learning methods focusing on handcrafted states, such as a fixed number with confidence coefficient of the generation amount, which limited the capability of dealing with virtual power plant bidding with uncertainty renewable energy generation amount. This paper uses the pretext tasks to pre-train the feature extractors, but omits the task headers in bidding task to keep more information, instead of compressing the information to a number. This paper also proposes an end-to-end method, which takes the historical data as the input and outputs the action. Leveraging feature extraction networks to discover the information of uncertainty energy in the end-to-end method, then using the deep reinforcement learning method for bidding action generation. Our method outperforms methods using regression model to predict the generation and put the prediction result to states with an average increasement of 37.57% per day.