Search

International audience; Uncertainty in the upcoming production of photo-voltaic (PV) plants is a challenge for grid operations and also a source of revenue loss for PV plant operators participating in electricity markets, since they have to pay penalties for the mismatch between contracted and actual productions. Improving PV predictability is an area of intense research. In real-world applications, forecasts are often needed for different time frames (horizon, update frequency, etc.) and are derived by dedicated models for each time frame (i.e. for day ahead and for intra-day trading). This can result in both different forecasted values corresponding to the same horizon and discontinuities among time-frames. In this paper we address this problem by proposing a novel seamless probabilistic forecasting approach able to cover multiple time frames. It is based on the Analog Ensemble (AnEn) model, however it is adapted to consider the most appropriate input for each horizon from a pool of available input data. It is designed to be able to start at any time of day, for any forecast horizon, making it well-suited for applications like continuous trading. It is easy to maintain as it adapts to the latest data and does not need regular retraining. We enhance short-term predictability by considering data from satellite images and in situ measurements. The proposed model has low complexity compared to benchmark models and is trivially parallelizable. It achieves performance comparable to state-of-the-art models developed specifically for the short term (i.e. up to 6 hours) and the day ahead. The evaluation was carried out on a real-world case comprising three PV plants in France, over a period of one year.

International audience; Short-term forecasts of generation or demand are required as inputs into several power system management functions. Forecast models are in general tuned to provide optimal accuracy and reliable estimations of associated uncertainty. In a second step, these forecasts are used as input in tools that perform various functions such as scheduling, reserves estimation, and trading in electricity markets. These functions often integrate algorithms that in turn optimize a value-related criterion, like cost, reliability, income, etc. The literature has shown that in some processes, like energy trading, this value may be optimized if a specific quantile forecast is selected rather than the forecast considered as most accurate. In this paper we propose a new data-driven approach in which the two steps of forecasting and decision-making are unified into one function that is optimized considering a single criterion, i.e. the value featured in the decision-making process. This approach allows us to bypass the use of specific forecasting models and could be extended to any decision-making process for which the inputs, outputs and objective functions are well defined. An intermediate approach is evaluated where a meta-optimization is applied to tune the forecast model as a function of the value it brings. This intermediate approach can also prove efficient, but does not allow us to bypass the forecasting models. The applications considered to evaluate the concept are photovoltaic and wind power generation trading in a day-ahead market, where we simultaneously optimize a forecasting model and a trading strategy, considering the final revenue on the electricity market as the objective function. For the meta-optimization, an Analog Ensemble (AnEn) model is used to forecast PV production, coupled with a Support Vector Regression (SVR) to forecast market prices. For the data-driven approach, we use an Extreme Learning Machine (ELM). As the objective function is highly complex and non-linear, the optimization is carried out by Particle Swarm Optimization (PSO).