A crucial aspect of the mid-term hydrothermal coordination problem is the nonlinear variation of hydro power with the water head, which can be computed through forebay and tailrace curves of the reservoirs. More accurate representations may also include turbine efficiency curves as a function of head and discharge, and penstock head losses as a function of turbined outflow. Based on large set of historical data for the operation of hydro plants, we propose a data-driven model to obtain such curves, which comprises: interpolation/extrapolation procedures to fill missing data and extreme regions of the curve; computation of average weekly values to avoid an optimistic behavior when using instantaneous curves for larger time intervals; an additive model with polynomial splines for curve fitting. The methodology was applied to the real Brazilian system, showing a significant improvement in the accuracy of the hydro production function, without sensitive changes on CPU time.