Nuclear Electric currently have complex computational models of how plants will behave under these conditions, which allows them to compute plant data (e.g., reactor temperatures) from given grid frequency data. One approach to damage assessment would require several years'-worth of real grid data to be fed into this model and the corresponding damage computed (via "cycle distributions" created by their damage experts). The results of this analysis would demonstrate one of three possibilities: the damage may be acceptable under all reasonable operating conditions; or it may be acceptable except in the case of an exceptional abrupt change in grid frequency (caused by power transmission line failure, or another power station suddenly going off-line, for instance), in which case some kind of backup supply (e.g., gas boilers) would be required; or it may simply be unacceptable.

However, their current model runs in approximately real time, making it inappropriate for such a large amount of data: our problem was to suggest alternative approaches. Specifically, we were asked the following questions:

- Can component damage be reliably estimated directly from cycle distributions of grid frequency? i.e., are there maps from frequency cycle distributions to plant parameter cycle distributions?

- Can a simple model of plant dynamics be used to assess the potential for such maps?

- What methods can be used to select representative samples of grid frequency behaviour?

- What weightings should be applied to the selections?

- Is it possible to construct a "cycle transform" (Fourier transform) which will capture the essential features of grid frequency and which can then be inverted to generate simulated frequency transients?

We did not consider this last question, other than to say "probably not".

We were supplied with data of the actual grid frequency measurements for the evening of 29/7/95, and the corresponding plant responses (obtained using Nuclear Electric's current computational model). A simplified nonlinear mathematical model of the plant was also provided.

Two main approaches were considered: statistical prediction and analytical modelling via a reduction of the simplified plant model.