Short-term load forecasting using dynamic neural networks

This paper presents forecasting of short-term electricity load using dynamic neural networks, DNN, and an assessment of the neural networks stability to ascertain continued reliability. A comparative study between three different neural network architectures is set up: feed forward, Elman and the radial basis neural networks. The performance and stability of each DNN is evaluated by means of an extensive simulation study using actual hourly load data. The neural networks weights are dynamically adapted. Stability for each of the three different networks is determined through Eigen values analysis. Evaluation of the networks is done in terms of estimation performance, stability and the difficulty in training a particular network. The results show that the radial basis neural network architecture performs better than the rest with overall mean average percentage forecasting error of 2.6%. Eigen value analysis also shows that it is more reliable as it remains stable as the input varies.