Objective & Background
The electric system is evolving towards a smart architecture, where Renewable Energy Sources (RES) production proliferates in form of photovoltaic (PV) and wind generators. A crucial issue in modern power systems consists in supporting this large-scale proliferation by mitigating its negative impact on the grid stability, system operation and control, making it necessary to optimally manage the energy production and consumption at short time scales. In this context, the very short-term forecast horizon (i.e., from minutes to 3-6 hours ahead) is gaining further attention as the Italian Regulatory Authority for Energy, Networks and Environment could also change the infra-day market by making it continuous, foreseeing an increase in the traded volumes and opening the possibility of increasing the profits in the energy market. The aim of this work is then to combine autoregressive methods and numerical weather modeling data to optimally forecast PV and wind production up to 6 hours ahead.
Method
The forecast system is based on a multivariate Auto-Regressive Integrated Moving Average (ARIMA), which uses the output of a short-term multi-model forecast system as exogenous variable. The application is tested on a wind farm located in the south of Italy and two PV plants in central Italy. Wind power data cover the period January 2017 – June 2018 with hourly temporal resolution, while PV data are available at 15 minutes increments from January 2018 to November 2018. The ARIMA PV forecasts for the next 6 hours are generated simulating operational conditions, updating every 15 minutes the input power data up to the previous quarter hour. The training period is set to 1 month discarding at each update the oldest data. The ARIMA wind power forecasts are configured in a similar way, using though a longer training period of 1 year to cope with the unavailability of sub-hourly time steps.
For wind, the multi-model power forecasts are obtained by applying an Analog Ensemble (AnEn) to post-process four different meteorological model runs. The multi-model forecasts are used as an exogenous input of the ARIMA model according to each corresponding time step. For PV, the multi-model is based on the combination of different AnEn configurations applied to the same four model runs used for wind.
In both applications the method is compared with persistence and, in the case of PV, with another method developed in RSE, based on the analysis and extrapolation of satellite images for the nowcasting of the solar irradiance , using then the AnEn to estimate the PV production of each plant.
Principal Findings
For each test case the ARIMA model is able to obtain the best results up to 2 hours ahead, while for longer time horizons the multi-model forecasts yields lower error scores than all other methods.
Conclusion
It is demonstrated that combining the multivariate ARIMA model with a multi-model short-term power forecast allows obtaining better results in the very short-term time frame compared with the reference methods.