Time series forecasting with feedforward neural networks trained using particle swarm optimizers for dynamic environments

Several studies have applied particle swarm optimization (PSO) algorithms to train neural networks (NNs) for time series forecasting and the results indicated good performance. These studies, however, assumed static environments, making the PSO trained NNs unsuitable for forecasting many real-world time series which are generated by non-stationary processes. This study formulates training of a NN forecaster as a dynamic optimization problem, to investigate the application of a dynamic PSO algorithm to train NNs in forecasting time series in non-stationary environments. For this purpose, a set of experiments were conducted on three simulated and seven real-life time series forecasting problems under four different dynamic scenarios. Results obtained are compared to the results of NNs trained using a standard PSO and resilient backpropagation (Rprop). The results show that the NNs trained using dynamic PSO algorithms outperform the NNs trained using PSO and Rprop. These findings highlight the potential of using dynamic PSO in training NNs for real-world forecasting applications.