In real, complex plants, a sensitivity analysis of the effects that variations in the plant inputs and design parameters have on the outputs is of great importance both from the point of view of productivity and of safety. To a first approximation, sensitivity analysis consists of estimating the partial derivatives of the outputs with respect to the varied quantities. These derivatives cannot be obtained on the real plant directly since the effects of all the involved variables are intermixed. Therefore, one has to resort to suitable computational models and algorithms.

A new neural network approach that aims at creating a differentiable copy of the plant is proposed. A feature of the method is that the data for network training are collected with the system in nominal operation: This represents, indeed, a fundamental constraint for all risky plants, for which unrestrained playing is definitely not recommended. The sensitivity coefficients (partial derivatives) thereby obtained are applied for the regulation of the reactivity of a simulated pressurized water reactor in response to changes in the electric load at the power grid, so as to maintain the average temperature of the water in the reactor core at a constant value.