Metaheuristics-Based Input Selection for Neural Networks: Application in Short-Term Load Forecasting

The scope of the present study is the assessment of a Particle Swarm Optimization (PSO) method for optimal input selection for Short-Term Load Forecasting (STLF) tasks. The test case involves the STLF of a high to medium voltage transformer, where a Feed-Forward Neural Network (FFNN) is trained and applied. An effective training process is crucial to the forecasting model's success. Low number of inputs can lead to low potential for generalization. On the contrary, a high number may lead to large training duration. Thus, a trade-off between the generalization and duration should be accomplished. Using the PSO based method, the most correlated inputs are selected, a fact that leads to low prediction errors.