Transfer functions and most controller design methods need linear models. We get one by approximating the nonlinear model close to a steady operating point.
Taylor series
For a function $f(x)$ near the steady state $\bar x$, keep only the first-order term:
$$f(x) \approx f(\bar x) + \left.\frac{df}{dx}\right|_{\bar x}(x - \bar x)$$
Deviation variables
Define $h' = h - \bar h$ and $q' = q_{in} - \bar q_{in}$. Deviation variables are zero at the steady state, which makes initial conditions zero and Laplace transforms simple.
Tank example
Linearising $\sqrt{h}$ about $\bar h$: $\sqrt{h} \approx \sqrt{\bar h} + \dfrac{1}{2\sqrt{\bar h}}h'$. Subtracting the steady-state balance gives
$$A\frac{dh'}{dt} = q' - \frac{c_v}{2\sqrt{\bar h}}\,h' \quad\Rightarrow\quad \tau\frac{dh'}{dt} + h' = K q'$$
with $K = 2\sqrt{\bar h}/c_v$ and $\tau = 2A\sqrt{\bar h}/c_v$. Both depend on the operating point: the tank responds more slowly and more strongly at high levels.
The linear model is good for small changes. A common rule of thumb is to trust it for changes of up to about 10–20 % around the operating point.