Deep Reinforcement Learning-Based Self-Scheduling Strategy for a CAES-PV System Using Accurate Sky Images-Based Forecasting

Compressed air energy storage (CAES) is a scalable and clean energy storage technology with great potential in renewables accommodation. From the point of view of the facility owner participating in the energy market, the profit of a CAES-PV system's coordinated operation is still at a notable risk. This paper addresses this problem by using a novel model-free deep reinforcement learning (DRL) method to optimize the CAES energy arbitrage in the presence of a sky images-based short-term solar irradiance forecasting model. To overcome the risk associated with the highly intermittent solar power productions, and thus efficient participation in an electricity market, a hybrid forecasting model based on 2-D convolutional neural networks (CNNs) and bidirectional long short-term memory (BLSTM) units is developed to capture high levels of abstractions in solar irradiance data, especially during cloudy days. Moreover, the thermodynamic characteristics of the CAES facility are considered to achieve more realistic real-time scheduling results. The comparative results based on a realistic-based case study demonstrate the effectiveness and applicability of the proposed framework compared to the state-of-the-art methods in the recent literature.

Automated summary

This paper proposes a novel model-free deep reinforcement learning method to optimize the energy arbitrage of a compressed air energy storage (CAES) system using a sky images-based short-term solar irradiance forecasting model. A hybrid forecasting model based on 2-D convolutional neural networks and bidirectional long short-term memory units is developed to overcome the risk associated with intermittent solar power productions. The thermodynamic characteristics of the CAES facility are also considered to achieve more realistic real-time scheduling results.