Neural networks for pattern-based short-term load forecasting: A comparative study

• Neural network models for pattern-based short-term load forecasting are proposed.
• Pattern-based approach simplifies forecasting time series with multiple seasonal cycles.
• MLP, RBFNN, GRNN, fuzzy counterpropagation NN and SOM were compared.
• GRNN and one-neuron perceptron learned locally gave best results in STLF.

In this work several univariate approaches for short-term load forecasting based on neural networks are proposed and compared. They include: multilayer perceptron, radial basis function neural network, generalized regression neural network, fuzzy counterpropagation neural networks, and self-organizing maps. A common feature of these methods is learning from patterns of the seasonal cycles of load time series. Patterns used as input and output variables simplify the forecasting problem by filtering out a trend and seasonal variations of periods longer than a daily one. Nonstationarity in mean and variance is also eliminated. In the simulation studies using real power system data the neural network forecasting methods were tested and compared with other popular forecasting methods such as ARIMA and exponential smoothing. The best results were achieved for generalized regression neural network and one-neuron perceptron learned locally.