Short-term load and wind power forecasting using neural network-based prediction intervals.

Area of Science:

• Electrical Engineering
• Computational Intelligence
• Energy Systems

Background:

• Modern electrical power systems are transitioning towards decentralization, increasing reliance on variable renewable energy sources like wind and solar.
• This shift introduces significant uncertainty, making accurate load forecasting crucial for effective power system management.
• Traditional point forecasting methods struggle to address the inherent uncertainties in power system operations.

Purpose of the Study:

• To develop a robust method for quantifying forecast uncertainties in power systems.
• To implement and extend a neural network (NN)-based approach for constructing prediction intervals (PIs).
• To propose a novel problem formulation for optimizing prediction interval generation.

Main Methods:

• Application and extension of the lower upper bound estimation (LUBE) method with NN models.
• Formulation of a constrained single-objective optimization problem derived from a multi-objective problem.
• Integration of Particle Swarm Optimization (PSO) with a mutation operator to solve the optimization problem.

Main Results:

• The proposed PSO-based LUBE method successfully constructs high-quality prediction intervals (PIs).
• Validation using electrical demand data from Singapore and New South Wales, and wind power generation from Capital Wind Farm.
• Demonstrated ability to provide accurate PIs for both load and wind power forecasts in a timely manner.

Conclusions:

• The developed NN-based LUBE method effectively quantifies uncertainties in power system forecasting.
• The novel problem formulation and PSO integration offer an efficient solution for generating reliable prediction intervals.
• This approach is valuable for managing increasingly complex and decentralized energy systems.