Design of artificial neural networks for short-term load forecasting. II. Multilayer feedforward networks for peak load and valley load forecasting

For pt.I see ibid., vol.138, no.5, p.407-13 (1991). In part I of the paper, a neural network with unsupervised learning was proposed to identify the day types and compute the hourly load pattern by averaging the load patterns of the same day type. In this part of the paper a neural network, commonly referred to as the multilayer feedforward network, is developed to forecast daily peak load and valley load. Unlike the self-organising feature maps in part I, the multilayer feedforward network is a neural net with supervised learning. The neural net is first trained using historical weather and load data. Then the trained neural net is applied to predict daily peak load and valley load. These peak and valley loads, when combined with the hourly load pattern, can yield the desired hourly loads. Results from short-term load forecasting of the Taiwan power system are given to demonstrate the effectiveness of the proposed neural networks