This paper presents an artificial neural network (ANN) approach for electric load forecasting. The ANN learns the relationship between past, current, and future temperatures and loads. It interpolates among load and temperature data in a training dataset to provide forecasts. The average absolute errors for one-hour and 24-hour ahead forecasts on actual utility data are 1.40% and 2.06%, respectively, which are significantly better than the 4.22% error of a current forecasting technique. The paper discusses traditional methods for load forecasting, including time series and regression approaches. Time series methods often fail to account for weather variables, leading to inaccurate predictions. Regression methods assume linear relationships between weather variables and load, but these relationships are not stationary and depend on spatio-temporal factors. The authors propose a hybrid approach combining time series and regression techniques using a layered perceptron ANN. The ANN can model non-linear relationships and adapt to new data without assuming a functional relationship between load and weather variables. The ANN is trained using the generalized delta rule, and its performance is evaluated based on error rates. The ANN was tested on hourly temperature and load data from the Seattle/Tacoma area. The results show that the ANN provides accurate forecasts for peak load, total load, and hourly load. The average errors for these cases were 2.04%, 1.68%, and 1.40%, respectively. The ANN's performance was compared with a current forecasting technique, showing significant improvements. The paper concludes that the ANN is suitable for interpolating load and temperature data to provide future load patterns. It is more flexible than traditional regression methods and can adapt to changing weather and load conditions. The ANN requires training data well spread in the feature space for accurate results. The training time varies depending on the case, but a trained ANN requires only a few milliseconds for inference. The authors acknowledge the support of the Puget Sound Power and Light Co., the National Science Foundation, and the Washington Technology Center.This paper presents an artificial neural network (ANN) approach for electric load forecasting. The ANN learns the relationship between past, current, and future temperatures and loads. It interpolates among load and temperature data in a training dataset to provide forecasts. The average absolute errors for one-hour and 24-hour ahead forecasts on actual utility data are 1.40% and 2.06%, respectively, which are significantly better than the 4.22% error of a current forecasting technique. The paper discusses traditional methods for load forecasting, including time series and regression approaches. Time series methods often fail to account for weather variables, leading to inaccurate predictions. Regression methods assume linear relationships between weather variables and load, but these relationships are not stationary and depend on spatio-temporal factors. The authors propose a hybrid approach combining time series and regression techniques using a layered perceptron ANN. The ANN can model non-linear relationships and adapt to new data without assuming a functional relationship between load and weather variables. The ANN is trained using the generalized delta rule, and its performance is evaluated based on error rates. The ANN was tested on hourly temperature and load data from the Seattle/Tacoma area. The results show that the ANN provides accurate forecasts for peak load, total load, and hourly load. The average errors for these cases were 2.04%, 1.68%, and 1.40%, respectively. The ANN's performance was compared with a current forecasting technique, showing significant improvements. The paper concludes that the ANN is suitable for interpolating load and temperature data to provide future load patterns. It is more flexible than traditional regression methods and can adapt to changing weather and load conditions. The ANN requires training data well spread in the feature space for accurate results. The training time varies depending on the case, but a trained ANN requires only a few milliseconds for inference. The authors acknowledge the support of the Puget Sound Power and Light Co., the National Science Foundation, and the Washington Technology Center.