Hybrid Energy Storage Control Based on Prediction and Deep Reinforcement Learning Compensation for Wind Power Smoothing

Aiming at the problem of output power fluctuations and uncertainty in wind power generation systems, a hybrid energy storage control method based on prediction and deep reinforcement learning (DRL) compensation was proposed. Firstly, a CNN-BiLSTM network was employed in predicting the wind power, and an adaptive moving average filtering algorithm was used to set the smooth output power reference curve in the future time, determining the total power reference values of the hybrid energy storage system (HESS). Then, low-pass filtering was employed in allocating the power of the HESS, and the deep deterministic policy gradient (DDPG) in the DRL was used to learn to compensate for the allocation values of the low-pass filtering, determining the new control targets of HESS. Finally, according to the control objective, a model predictive control was employed in controlling the charging and discharging of the HESS, so as to smooth the wind power and ensure that the state of charge of the HESS was maintained within a reasonable range. A simulation experiment was conducted using the data of a real wind farm and the experimental results show that the grid-connected power fluctuations are effectively reduced, the overcharge and over-discharge of HESS are prevented, and the working life of HESS is extended with the proposed strategy.