Forecasting Electricity Demand with Dynamic Characteristics Based on Signal Analysis and Machine Learning

In daily power supply, it is important to accurately forecast the electricity demand for each time of the next day. Electric demand at a certain time is affected not only by weather conditions at that time but also by weather conditions and electric power demand in the past several hours. The dynamic characteristics of electricity demand are nonlinear and vary with time of day, day of week, and season. We focused on the frequency spectrum of the electricity demand curve as a feature of the electricity demand with dynamic characteristics which change with time. We propose a method of pattern classification of dynamic characteristics by clustering frequency spectrum, and a method of forecasting frequency spectrum by nonlinear regression based on kernel method. The former is an approach that emphasizes explanatory, and the latter is an approach that emphasizes forecast accuracy. We confirmed the usefulness by the experiment using the open data.