Near and long-term load prediction using radial basis function networks

Conference · Sun Dec 31 04:00:00 EST 1995

OSTI ID:377057

Hancock, M F ^[1]

---

1. Rollins College, Winter Park, FL (United States)

A number of researchers have investigated the application of multi-layer perceptrons (MLP`s), a variety of neural network, to the problem of short-term load forecasting for electric utilities (e.g., Rahman & Hazin, IEEE Trans. Power Systems, May 1993). {open_quotes}Short-term{close_quotes} in this context typically means {open_quotes}next day{close_quotes}. These forecasts have been based upon previous day actual loads and meteorological factors (e.g., max-min temperature, relative humidity). We describe the application of radial basis function networks (RBF`s) to the {open_quotes}long-term{close_quotes} (next year) load forecasting problem. The RBF network performs a two-stage classification based upon annual average loads and meteorological data. During stage 1, discrete classification is performed using radius-limited elements. During stage 2, a multi-layer perceptron may be applied. The quantized output is used to correct a prediction template. The stage 1 classifier is trained by maximizing an objective function (the {open_quotes}disambiguity{close_quotes}). The stage 2 MLP`s are trained by standard back-propagation. This work uses 12 months of hourly meteorological data, and the corresponding hourly load data for both commercial and residential feeders. At the current stage of development, the RBF machine can train on 20% of the weather/load data (selected by simple linear sampling), and estimate the hourly load for an entire year (8,760 data points) with 9.1% error (RMS, relative to daily peak load). (By comparison, monthly mean profiles perform at c. 12% error.) The best short-term load forecasters operate in the 2% error range. The current system is an engineering prototype, and development is continuing.

• Research Organization: Pacific Northwest Lab., Richland, WA (United States)
• OSTI ID: 377057
• Report Number(s): PNL-SA--26375; CONF-9503142--; ON: DE96009360
• Country of Publication: United States
• Language: English

Related Subjects

29 ENERGY PLANNING, POLICY, AND ECONOMY
99 GENERAL AND MISCELLANEOUS
ELECTRIC UTILITIES
FORECASTING
LOAD ANALYSIS
NEURAL NETWORKS