Combine Particle Swarm Optimization with Artificial Neural Networks for Short-Term Load Forecasting

Abstract

Electricity consumption curves are highly non-linear as many external factors affect the electricity consumption. Artificial Neural Networks (ANNs) are popular in electricity load forecasting since its pattern recognizing and learning abilities of data. Training ANNs are important as it directly affects the forecasts. However, backpropagation training algorithm likely to stops at local minima. Therefore, this research uses Particle Swarm Optimization to train an ANN for forecasting short-term load demand in Thailand. One-year training data is used to forecast the days in 2013. Forecast are evaluated in terms of the Mean Absolute Percentage Error (MAPE). Monthly, yearly, weekdays’, Mondays’, weekends’, holidays’, and bridging holidays’ average MAPEs from PSO are compared with MAPEs from backpropagation training algorithm. Average MSPEs show that PSO outperforms backpropagation for training ANNs in short-term load forecasting.