Wireless Sensor Networks (WSN) are networks composed by tiny devices equipped with sensing, processing, storage, and wireless communication capabilities. WSN are used in highly dynamic environments. Applications for WSN should have an autonomous behavior to adapt their operation and achieve the best network performance. Such adaptation should preferably be performed by a middleware layer tailored to the limited resources of WSN. In this paper, we introduce a Reference Architecture (RA) of a self-adaptive middleware for WSN to contribute for the development of solutions enabling autonomic behavior in WSN. Our RA follows an autonomic computing model (MAPE-K) proposed by IBM and it was specified using a formal description language (pi-ADL) that enables the specification of dynamic architectures. ProSA-RA was used to systematize the design, representation and evaluation of our RA.