In "I take 3072 for Paillier's $n$", 3072 is surely the bit size of $n$. Thus I'll read the question as:
How wide should be OU's $n=p^2q$ to be as safe as Paillier's $n=pq$ of 3072 bits?
The best known attack against both cryptosystems is the factorization of $n$.
The best known factorization method for $n=pq$ with $p$ and $q$ random primes of roughly the same size is GNFS, of cost $L_n[1/3,4\cdot3^{-2/3}]$, per L notation.
For factorization of $n=p^2q$, GNFS also works with similar cost, thus we must have $n$ at least 3072-bit. However, Lenstra's ECM is also to consider, and (I think) it's cost is about $L_{\min(p,q)}[1/2,2^{1/2}]$. Thus to maximize resistance to that later algorithm we should have $p$ and $q$ of next to the same size. That size must be at least 1024-bit to get a 3072-bit $n$. And if we do the math and ignore the $o(1)$ in $L_k[\alpha,c]=e^{(c+o(1))\ln(k)^\alpha\ln(\ln(k))^{1-\alpha}}$, we get that 1024-bit $p$ and $q$ is (barely) enough for ECM to be more costly than GNFS.
Hence we should have $p$ and $q$ of 1024-bit at least, e.g. in range $[2^{1024-1/3},2^{1024}]$ for 3072-bit $n$. If we want to err on the safe side because we ignored the $o(1)$, we can bump that a little to e.g. 1152-bit, e.g. in range $[2^{1152-1/3},2^{1152}]$ for 3456-bit $n$.
The same computation supports the mysterious Captain Nemo's "RSA Moduli Should Have 3 Prime Factors" quote.
Addition: supporting evidence in Mathematica, yielding 0.5… (resp. 10.3…) for the base-2 logarithm of the ratio of work between ECM @1024-bit (resp. @1152 bits) factors, to the work for GNFS @3072-bit product. Try It Online!.
L[n_, a_, c_] := Exp[c (Log[n]^a) (Log[Log[n]]^(1-a))];
LGNFS[n_] := L[n, 1/3, 4 3^(-2/3)];
LLenstra[p_] := L[p, 1/2, 2^(1/2)];
Log[2., LLenstra[2^{1024,1152}]/LGNFS[2^3072]]
