# Compute the gradient for the previous time step
dh_prev = dh * (1 - z) + np.dot(self.wz[:, :self.hidden_size].T, dz) + np.dot(self.wr[:, :self.hidden_size].T, dr)