Hyperspectral image fusion (HIF) is critical to a wide range of applications in remote sensing and many computer vision applications. Most traditional HIF methods assume that the observation model is predefined or known. However, in real applications, the observation model involved are often complicated and unknown, which leads to the serious performance drop of many advanced HIF methods. Also, deep learning methods can achieve outstanding performance, but they generally require a large number of image pairs for model training, which are difficult to obtain in realistic scenarios. Towards these issues, we proposed a self-regression learning method that alternatively reconstructs hyperspectral image (HSI) and estimate the observation model. In particular, we adopt an invertible neural network (INN) for restoring the HSI, and two fully-connected network (FCN) for estimating the observation model. Moreover, \emph{SoftMax} nonlinearity is applied to the FCN for satisfying the non-negative, sparsity and equality constraints. Besides, we proposed a local consistency loss function to constrain the observation model by exploring domain specific knowledge. Finally, we proposed an angular loss function to improve spectral reconstruction accuracy. Extensive experiments on both synthetic and real-world dataset show that our model can outperform the state-of-the-art methods