This paper presents a filter-based algorithm called PREDICTOR for optimizing multidimensional K-NN queries in WSN. A filter installed at each sensor node is a node value distribution range. It is used to prevent the node from sending the data that belongs to the covering range of the filter, and so the node's energy is saved. The server keeps the historical sample data of all the nodes, and determines filters for them according to the query requirement and the samples. Three optimization strategies are proposed: (1) the method for dynamically adjusting the assignment strategies of the filter covering ranges so that a node with little chance to contribute to the final result is assigned with a larger covering range; (2) the method for sharing the filter among nodes so that the nodes with similar historical data are assigned with the same filter; (3) the method for transmitting filters in a compressed way to reduce the cost of filter updating for different K-NN queries. Evaluation experiment results prove the efficiency of the algorithm PREDICTOR in energy saving. Compared with the naive method, this approach can reduce the transmission volume dramatically.