Abstract:

With the construction of a new power system dominated by new energy sources, the uncertainty of power load has increased significantly, posing severe challenges to the safe, stable and economic operation of power grids. Against this background, uncertain load forecasting methods such as probabilistic forecasting and interval forecasting have attracted extensive attention. Scenario technology provides key inputs for forecasting models by modeling and simulating multi-source uncertainties including load, meteorology and new energy output. This paper proposes a short-term load forecasting scenario generation method that integrates deep reinforcement learning(DRL) and conditional diffusion model(CD). Aiming at the complex coupling and dynamic characteristics of multivariate time-series data such as load and meteorology, a conditional diffusion model combined with bidirectional long short-term memory(Bi-LST) network, self-attention mechanism and seasonal decomposition layer is designed to accurately leamn the intrinsic conditional probability distribution of data and generate high-fidelity fiture scenarios. Meanwhile, to address the difficulty of hyperparameter tuning, an optimization framework based on DRL, is constructed, which formulates hyperparameter optimization as a lMarkov decision process and realizes adaptive parameter configuration through the interaction between agents and the environment. Experiments based on actual load and meteorological lata from a regton in China show that the proposed method outperforms benchmark models in various evaluation indicators.

Key words: short-term load forecasting, scenario generation, conditional diffusion model, deep reinforcement learning, uncertainty quantification