Power market clearing price prediction and confidence interval estimation with fast neural network learning

Market clearing prices (MCPs) play an important role in a deregulated power market, and good MCP prediction and interval estimation will help utilities and independent power producers submit effective bids with low risks in this uncertain market. Since MCP is a nonstationary process, an adaptive algorithm with fast convergence is important. A common method for MCP prediction is neural networks, and multilayer perceptron networks (MLP) is one of the widely used networks. Backpropagation (BP) is a popular learning method for MLP, while BP suffers from slow convergence. This paper presents an integrated learning and interval estimation algorithm for MCP prediction. In the extended Kalman filter (EKF) framework, confidence interval is a natural by-product of EKF, and is integrated with learning process to improve learning results in addition to fast convergence. Since Kalman filter (KF) is a minimum variance estimator for linear system, EKF framework helps to provide a smaller confidence interval, which is preferred in risk management. Testing results on New England MCP prediction show the integrated learning and confidence interval algorithm provides better prediction than BP algorithm and the confidence interval is smaller with reasonable coverage than a Bayesian inference-based interval estimation method.