Objective & Background
A microgrid is a small-scale power system comprising loads, distributed energy resources (photovoltaic, wind turbines, conventional generators) and energy storage systems. The energy management system (EMS) establish an optimal resources planning according to operational objectives. This optimization layer is generally done at a sub-hourly time frame. Since the weather determines the renewable energy sources generation and load, EMS needs weather forecasts. This presentation gives a methodology to get weather forecast models for sub-hourly data frequency through a time series approach. Weather variables considered are irradiance, temperature, wind speed, humidity rate and air pressure from Singapore at a 15min time step.
Multiple seasonal trigonometric polynomials are used to fit the unconditional mean and variance of the weather variables time series. Intra-year and intra-day seasonality are considered as well as their multiplicative interactions. The multiple seasonal trigonometric polynomials orders are selected by Schwarz’s Bayesian criterion through grid searches on trigonometric polynomials and autoregressive orders of the time series (to whiten the estimated residuals). An autoregressive model is fitted on the resulting seasonally centered and scaled weather variable. The autoregressive order is also selected by Schwarz’s Bayesian criterion. More than 5 years of data are used for the training dataset and a full year is used for the test dataset. The forecast horizon ranges from +15min to +6h.
Forecast accuracy of the model is benchmarked against persistence, anomaly persistence and standardized persistence forecast accuracies. It shows that each modeling step contributes to increase the accuracy of the forecast compared to persistence even at the very short term horizon.
To go further in the analysis of sub-hourly weather time series the next steps will involve vector autoregressions and multiple weather regimes.