Gene-expression programming—an electrical-load forecast alternative to artificial neural networks

We introduce a method called gene-expression programming (GEP, a variant of genetic programming) to estimate electrical load via a nonlinear combination of ensemble NWP forecasts. From a population of competing and evolving algorithms (each of which can uses a different combination of NWP ensemble members), GEP uses computational natural selection to find the algorithm that maximizes a verification fitness function for electrical load. For western Canada, the typical electrical load curve has relative minima during nighttime and mid-day, and relative maxima in the morning and evening. We estimate the load for each of these four extrema by first removing the trend, annual cycle, and semi-annual cycle as found via Fourier analysis. Then, we use GEP to fit the remaining load signal (separately for each extreme) as a function of NWP forecasts of air temperature, wind speed, precipitation, humidity, day-of-the-week, and other meteorological and calendar variables. This is done via a multi-year training set of load observations and archived multi-model ensemble forecasts. Once these best-fit algorithms are found, we use them to make real-time forecasts for each of the load extrema, based on real-time NWP forecasts. We then use a Bezier curve to interpolate between the extrema to get hourly load forecasts. Details will be presented, along with verification results for independent data set, and comparisons with some other load-forecasting methods.