login name
ITA english

Report RSE 16001370

Modelli di errore di previsione della generazione rinnovabile e del carico, per funzioni di analisi in linea della sicurezza


First of all, the report presents the statistical analysis of historical data (both forecast and snapshots of power exchanges) to evaluate the dependence of forecast errors from forecasted values. The division of historical data into clusters characterized by similar prediction time horizons has allowed to analyze the dependence of forecast accuracy on the forecast time horizon: the means and standard deviations of forecast errors for the variables are compared on different clusters and the results show that there is not a clear trend between time horizon and forecast accuracy: in some cases, the means and the standard deviation of the errors are reduced with decreasing time horizon (expected behavior), but in other cases they increase when reducing forecast horizon, or are scarcely influenced. The comparison of the probability distributions of the forecast errors for different cluster of horizons confirms the abovementioned results.

In general, defining the uncertainty cloud of the grid states around a forecast operating condition requires the knowledge of the “forecast state”. In other words, the Monte Carlo extraction of the samples of the stochastic injections has to account for the forecast values of the injections themselves.

The report presents an overview of possible methods to extract the samples of the injections conditioned to the relevant forecast values, starting from the method of "nearest neighbor" up to Gibbs sampling, to the conditional sampling based on Nataf transform and to the estimator based on kernel functions. The simulations run on a simple test network with few random variables have shown how the sampling based on Nataf transform represents the best compromise between computational speed (fundamental for on-line applications) and accuracy.

The report presents in detail the problems and the solutions adopted for the implementation of the Nataf transform based conditional sampling method within the on-line platform of the European project iTesla for security assessment. The application of the sampling method to extended data sets (containing over 6000 variables and relative to the French network) confirms the compatibility of the calculation times with "online" requirements. The analysis shows that the forecast values of the injections are often outside the cloud of extracted samples for the same injections; this is due to the poor accuracy of the forecasts found in the historical data and used to build the uncertainty model. The analysis has also allowed to list the variables that have an increased difference between the forecast value and the corresponding set of extracted samples.

Finally, the report presents the methodology for the validation of the on-line iTesla platform for security assessment: in particular, we highlight the changes to be made to the main workflow of the platform for validation purposes. In addition the report describes the procedures and the metrics to define the accuracy and computational efficiency of the individual modules of the on-line platform.

Related Links