Hourly load forecasting by neural networks
- Southern Methodist Univ., Dallas, TX (United States)
- Electric Power Research Institute, Palo Alto, CA (United States)
In this work, a new method based on neural network (NN) technology for forecasting of hourly electric system loads is developed. The forecaster consists of three modules each consisting of several multi-layered feed forward NNs within it. These modules model the weekly, daily, and hourly trends of load-weather relationship. The final forecast is obtained by combining individual predictions made by each module through an adaptive filtering scheme. A modified version of {open_quotes}error back-propagation{close_quotes} learning rule is developed in this work which adjust NN parameters adaptively during on-line forecasting. Average hourly prediction errors obtained in extensive off-line studies with data from four different utilities of various sizes are around 2% for one-day-ahead forecasts and below 3% for two-to-five-day-ahead predictions. The package performs well in cases of rapid changes in temperature (below 3% error for temperature swings as large as 20{degrees}F). The performance is compared to several other forecasting techniques with favorable results. The forecaster has already been implemeted on-line at two utilities and is scheduled to be implemented at six other utilities in Fall 1993.
- Research Organization:
- Electric Power Research Inst., Palo Alto, CA (United States); Pacific Consulting Services, Albany, CA (United States)
- OSTI ID:
- 103288
- Report Number(s):
- EPRI-TR--105012; CONF-930969--; CNN: Contract RP2473-44; Contract 03613-004
- Country of Publication:
- United States
- Language:
- English
Similar Records
An adaptive modular artificial neural network hourly load forecaster and its implementation at electric utilities
An artificial neutral network hourly temperature forecaster with applications in load forecasting