Resumen
Assessing the training process of artificial neural networks (ANNs) is vital for enhancing their performance and broadening their applicability. This paper employs the Monte Carlo simulation (MCS) technique, integrated with a stopping criterion, to construct the probability distribution of the learning error of an ANN designed for short-term forecasting. The training and validation processes were conducted multiple times, each time considering a unique random starting point, and the subsequent forecasting error was calculated one step ahead. From this, we ascertained the probability of having obtained all the local optima. Our extensive computational analysis involved training a shallow feedforward neural network (FFNN) using wind power and load demand data from the transmission systems of the Netherlands and Germany. Furthermore, the analysis was expanded to include wind speed prediction using a long short-term memory (LSTM) network at a site in Spain. The improvement gained from the FFNN, which has a high probability of being the global optimum, ranges from 0.7% to 8.6%, depending on the forecasting variable. This solution outperforms the persistent model by between 5.5% and 20.3%. For wind speed predictions using an LSTM, the improvement over an average-trained network stands at 9.5%, and is 6% superior to the persistent approach. These outcomes suggest that the advantages of exhaustive search vary based on the problem being analyzed and the type of network in use. The MCS method we implemented, which estimates the probability of identifying all local optima, can act as a foundational step for other techniques like Bayesian model selection, which assumes that the global optimum is encompassed within the available hypotheses.