Electric Power Prediction Model based on Bayesian Deep Q-network

Electric power forecasting is important for the stable operation and efficiency optimization of power grid. This paper proposes a Bayesian Deep Q-Network (BDQN) based power prediction model, which combines the uncertainty estimation ability of Bayesian neural network with the sequential decision ability of deep reinforcement learning. A new feature extraction method for time series and an adaptive reward function are designed to balance the prediction accuracy and model stability. Experiments show that the proposed BDQN model has obvious advantages in forecasting accuracy and quantification of uncertainty, and can provide a new way for power forecasting and related management decision-making.