Abstract: The paper presents a novel approach for hourly short term forecast of load active power using discrete state Hidden Markov Models. The load data used belongs to the New York Independent System Operator (NYISO) and has been recorded in the years 2014-2017. In the first step, features the best explaining load power changes from the set of weather data, market data (price for load, losses or congestion) and calendar data (day type, day of week, season) are defined. Due to strong seasonality in the data, also a filtering step is included. Finally, the forecast itself is executed with 24 discrete state Hidden Markov Models with a high number of states. Besides the direct comparison with the forecast results obtained by NYISO, the approach is evaluated using an additional benchmark method.
Keywords: Short term load forecast, discrete state Hidden Markov
Model, Gaussian mixture model, multivariate normal distribution, clustering, filtering, Euclidean distance
