A Weather-Based Forecasting Method for Short-Term Aggregate Power Loads

This paper introduces a weather-based method for short-term forecasting of aggregate electricity load. We extract the weatherand illumination-dependent load via least-squares fitting for load versus Steadman apparent temperature and logscale natural illumination. A separate fit is done for each hour of the day, and then Fourier-transform-based spectral analysis handles the resulting residual. The input consists of past load and weather data (temperature, humidity and cloud cover), weather forecasts, and location information. We do extensive performance tests using load data from a mid-size U.S. city as well as for 8 climate zones in the state of Texas. A well-tested exponential smoothing method due to J.W. Taylor serves as a benchmark. The results show that the benchmark outperforms our method up to about four hours ahead, but the new technique is considerably better for longer lead times. The method is highly robust and yields accurate forecasts for the next day or as far ahead as the weather can be forecasted. These characteristics make it a good forecasting tool for supporting decisions for unit commitment, economic dispatching, and electricity purchase in day-ahead markets.