The report describes the activities carried out in synergy with the FP7 European project iTesla and within the subsequent “iTesla Power System Tools” open source project. The iTesla project has developed a methodology and a software platform for the evaluation of the dynamic security of the electrical system for in-line applications, and up to a few hours ahead of the real time, taking into account the uncertainties in the forecasting errors related to the power absorption of loads and power injections from non dispatchable renewable energy plants.

The first part of the report recalls the methodology of validation for the in-line platform modules, with particular attention to the MCLA (Monte Carlo Like Approach) module. In particular, the validation aims to verify the ability of decision trees used in the MCLA module to quickly and correctly assess the security of states belonging to the cloud of uncertain states around the expected system operating condition. The application of the MCLA module to a use case in the French EHV transmission grid shows the appropriateness of the decision trees (DT) performance; as far as the analysed contingencies and security issues are concerned, the DTs ensure low rates of false alarms and very low rates of missed alarms. The critical issues identified in terms of false and/or missed alarms are due to non-optimal choice of the set of historical data for the training of trees and/or possible changes of network topology between the set of states of the historical states and the forecast state specifically analysed with the inline MCLA module.

The validation led to identify further research needs, namely (1) a thorough analysis of the quality of historical data with a focus on network topology, with the aim to improve the forecast uncertainty model, and (2) a verification of the accuracy of the forecast error model conditioned to the specific forecast state.

The statistical analysis of historical data has identified few variables with high variance; many variables have a distribution strongly sensitive to the presence of outliers: some of them strongly differ from the rest of the realisations of the variable, so they alter the resulting distribution. Some outliers are very numerous, generating peaks in the distribution that are completely removed when they are extracted, and the relevant distribution passes from bimodal to unimodal. Finally, quite a few variables have an average value significantly different from zero: most of them present a multimodal distribution (i.e. they exhibit more than one peak in the probability density curve).

The report also describes the outcome of a preliminary “ex-post” test performed to check a necessary though not sufficient condition for the accuracy of the conditioned uncertainty model: for each forecast state, the actual snapshot of each variable should lie within the confidence interval of the same variable sampled with the conditioned uncertainty model in the MCLA module. The tests provide encouraging results and useful information to refine the accuracy of the model: in particular, the variables with the snapshot more frequently outside the confidence interval constitute a relatively small group compared to the total number of variables.