Agent-based short-term load and price forecasting using a parallel implementation of an adaptive PSO-trained local linear wavelet neural network

Short-Term Load and Price forecasting are crucial to the stability of electricity markets and to the profitability of the involved parties. The work presented here makes use of a Local Linear Wavelet Neural Network (LLWNN) trained by a special adaptive version of the Particle Swarm Optimization algorithm and implemented as parallel process in CUDA. Experiments for short term load and price forecasting, up to 24 hours ahead, were conducted for energy market datasets from Greece and the USA. In addition, the fast response time of the system enabled its encapsulation in a PowerTAC agent, competing in a real time environment. The system displayed robust all-around performance in a plethora of real and simulated energy markets, each characterized by unique patterns and deviations. The low forecasting error, real time performance and the significant increase in the profitability of an energy market agent show that our approach is a powerful prediction tool, with multiple expansion possibilities.