Combining the Transformer Model with DDPG to Improve the Dynamic Scheduling Efficiency of Photovoltaic Power Station Intelligent Grid Connection

In view of the large power fluctuation (daily fluctuation rate>35%), the dynamic response lag of traditional scheduling methods (average delay>120s), and the lack of multi-time scale coordination capabilities in the process of photovoltaic power station grid connection, this study proposes an intelligent scheduling framework that integrates Transformer time series modeling and deep deterministic policy gradient (DDPG). By constructing a Markov decision process with 12-dimensional state vectors including photovoltaic output, energy storage SOC, and grid frequency deviation, a spatiotemporal feature extraction method based on a multi-head self-attention mechanism (8 heads, hidden layer dimension 512) is designed. The module encodes the historical power data at 15-minute intervals (sliding window length T=96), generates a 128-dimensional latent state vector with time series correlation and inputs it into the Actor network of DDPG. Combined with the Critic network, a multi-objective reward function is constructed, which includes grid stability reward (frequency deviation Δf<0.2Hz), energy storage loss penalty (weight λ=0.8 when charging and discharging depth DOD>80%) and equipment regulation cost (limit on the number of inverter switching). The training is carried out using the priority experience replay mechanism and the double-delay update strategy (update period τ=100). In the simulation environment of a 10MW photovoltaic power station, the model shortens the dispatch response time to 36.2 seconds and reduces the power fluctuation rate to below 9.1%, verifying the Transformer-DDPG framework’s ability to capture complex spatiotemporal correlation features and its multi-objective optimization advantages in dynamic scheduling tasks, and providing an intelligent solution for the grid connection of high-penetration renewable energy.