抄録
A Multi-layered-type neural network is an attractive technique for daily electric load forecasting because the neural network can acquire nonlinear relationship among the electric load data and their factors(weather, temperature, etc.) automatically. In this paper, we first discuss some essential issues to be considered in neural network applications. One is difficulty of getting sufficient effective training data, and one is influence of abnormal learning data, and the other is inevitable outerpolation. For these issues, we developed following three methods in order to forecast more accurately (1) a structure of the neural networks for insufficient training data, (2) detection and diminishing the influence of abnormal data, (3) employment of interpolation network and outerpolation network with additional data for outerpolation. Furthermore, to raise sensitivity between electric loads and factors, we developed (4) removal of base load. Those methods effectively work to decrease the average absolute errors of peak-load forecasting and 24-hour load forecasting to 1.78% and 2.73% respectively.
