Twin delayed deep deterministic policy gradient based stochastic optimal dispatch of hydro and wind power considering the uncertainty of wind power in power-time dimensions

The randomness, fluctuation and uncertainty of wind power brings great challenges to the dispatch and control of the power system. In order to raise the economy of hydro and wind power optimal dispatch, a stochastic optimal dispatching model considering the uncertainty of wind power in both time and power dimensions is proposed. The versatile distribution is used to describe the uncertainty of wind power in different power-time intervals, providing foundation for positive and negative reserves of hydro power. Moreover, to solve the multi-objective stochastic optimal dispatching model, the twin delayed deep deterministic policy gradient (TD3) algorithm is utilized, which can avoid falling into local optimum and has stronger search ability when dealing with high-dimensional multi-objective optimization problem. Simulations in a wind farm and hydro power station of western China show that the proposed model can accurately describe the uncertainty of wind power, and TD3 algorithm can find the global optimal solution more effectively than the traditional intelligent algorithms and other deep reinforcement learning algorithms. • In hydro and wind power dispatch, a two-dimensional time-power interval is proposed to segment wind power, employing a versatile distribution model to enhance the accuracy of uncertainty characterization. Additionally, a similarity index based on KL divergence clustering is introduced to quantify wind power interval uncertainty, simplifying the complexity of uncertainty patterns in the dispatch model. This approach reduces the conservativeness of dispatch decisions by streamlining uncertainty representation, ultimately improving the efficiency and flexibility of hydro-wind dispatch strategies. • A stochastic optimization strategy for hydro and wind dispatch based on the TD3 algorithm is presented, capable of directly solving dispatch planning problems involving nonlinear objective functions and constraints. This strategy uses Markov decision processes for decision-making, making it less prone to local optima compablack to existing algorithms.