서강대학교

초록 (요약문)

Recently, renewable energy is rapidly integrated into the power grid to prevent climate change, and accurate forecasting of renewable generation becomes critical for reliable power system operation. By leveraging deep learning on energy big data, forecasting techniques have been actively developed to reduce the uncertainty of renewable generation. However, reducing the volatility of renewable generation also contributes to stabilizing the power grid in addition to improving forecast accuracy. In doing this, distributed renewable energy sources can be aggregated to stabilize the output variability. Also, large-scale energy storages such as lithium-ion batteries are used to absorb fluctuations of renewable generation. In this dissertation, we propose novel forecasting strategies considering volatility mitigation as well as forecasting accuracy improvement. First, we propose a novel short-term forecasting technique called space-time convolutional neural network (STCNN).We construct a space-time matrix to exploit the location information of multiple renewable energy sites and historical generation data, which is learned by a convolutional neural network. The proposed multi-site forecasting framework is simple but effective for both individual site forecasting and aggregated generation forecasting. Second, we propose a novel strategy called error compensable forecasting. We switch the objective of forecasting from reducing errors to making errors compensable by leveraging a battery. The challenging part of the proposed objective lies in that the stored energy at current time is affected by the previous forecasting result. In this regard, we propose a deep reinforcement learning (DRL) based error compensable forecasting framework, called DeepComp, having forecasting in the loop of control. Also, we leverage proximal policy optimization, which is simple to implement with outstanding performance for continuous control. Finally, we extend the proposed DeepComp into the bidding in the real-time renewable energy market, where renewable producers can maximize their profits from the generation bids by leveraging large-scale batteries. However, most studies focus on either increasing revenues by energy arbitrage or reducing deviation penalties by renewable generation error compensation. In this regard, we propose a DRL based real-time renewable energy bidding strategy, called DeepBid to improve bidding performance by increasing revenues and reducing deviation penalties simultaneously. The proposed DeepBid strategy substantially increases the total profit compared to existing bidding strategies by achieving as high revenues as the arbitrage strategy and as low deviation penalties as the error compensation strategy.