Abstract: A novel hybrid simulation method for a deep deterministic policy gradient (DDPG) algorithm and generalized S-transform (GST), referred to as DDPG-GST, is pro-posed. In the DDPG-GST method, empirical mode decomposition is first used to decompose the original data into nonstationary fluctuating wind speed components and trend components. The GST is then used to extract the time–frequency characteristics of the nonstationary fluctuating wind speed components, followed by the construction of the GST time–frequency power spectrum matrix. Subsequently, Cholesky decomposition is applied to generate simulated nonstationary fluctuating wind speeds. These simulated speeds are input into the DDPG network for regulation and control to optimize the simulation pro-cess. Finally, the simulated total wind speeds are obtained by superposing the simulated nonstationary fluctuating wind speeds with the trend components. The results show that DDPG-GST retains the energy characteristics of nonstationary fluctuating wind speeds more accurately in the time domain compared to the GST simulation method. Additionally,the energy distributions, derived from the GST coefficient amplitudes by the DDPG-GST method in the time-frequency domain, align more closely with the targets. The average power spectrum of the DDPG-GST method is closer to the target. Therefore, the non-stationary wind speed simulation based on deep reinforcement learning is a high-precision, data-driven simulation method.

Key words: non-stationary wind speed simulation, deep reinforcement learning, S-transform, regulation and control