Development of feed-forward neural network models for gas short-term load forecasting
- Marquette Univ., Milwaukee, WI (United States)
The development of feed-forward artificial neural network based models to predict gas consumption on a daily basis is the subject of this paper. The discussion concerns an iterative process based on network sensitivities and intuition regarding selection of proper input factors. The method is applied in forecasting gas consumption for two regions in the State of Wisconsin, namely, a portion of metropolitan Milwaukee and a region near Fond du Lac. The investigation includes a study of the effects of various network sizes and training algorithms given a limited availability of relevant historical data. The effects of using multiple sources of weather data are also investigated for regions without a centrally located weather recording station. The training results indicate that feed-forward artificial neural network based models reduce the residual predicted consumption root mean squared errors by more than half when compared to models based on linear regression using identical input factors. The inclusion of weather data from multiple sources leads to further error reduction. The models were then used for load forecasting starting in Summer 1994. The static models performed well from the Summer 1993 until the severe cold weather of mid-January 1994. This poor performance was scrutinized; investigations point to probable causes that include severe weather patterns non-existent in the training data and the excitation of consumer behavioral modes either not evident in the training data or not well represented by time-of-the-year input factor proxies. The search for a solution lead to dynamic models that were implemented by updating the neural networks with new data as it became available. These models performed significantly better but still poorly during and after the severe weather of January 1994.
- Research Organization:
- USDOE Pittsburgh Energy Technology Center (PETC), PA (United States); Oregon State Univ., Corvallis, OR (United States). Dept. of Computer Science; Naval Research Lab., Washington, DC (United States); Electric Power Research Inst. (EPRI), Palo Alto, CA (United States); Bureau of Mines, Washington, DC (United States)
- OSTI ID:
- 81604
- Report Number(s):
- CONF-9410335-; ON: DE95011702; TRN: 95:004731-0019
- Resource Relation:
- Conference: Adaptive control systems technology symposium, Pittsburgh, PA (United States), 24-25 Oct 1994; Other Information: PBD: [1994]; Related Information: Is Part Of High-tech controls for energy and environment. Proceedings; Biondo, S.J.; Drummond, C.J. [eds.] [USDOE Pittsburgh Energy Technology Center, PA (United States)]; PB: 287 p.
- Country of Publication:
- United States
- Language:
- English
Similar Records
An adaptive modular artificial neural network hourly load forecaster and its implementation at electric utilities
Predicting effects of climate fluctuations for water management by applying neural network