A load forecasting technique based on a combinedmodel formed by ensemble Kalman filter（EnKF）and phase -spacereconstruction（PSR）is proposed to optimize forecasting results.For the load data time-series, phase-space reconstruction is firstlyimplemented using delayed coordinate embedding. Future loadstates can be predicted by local averaging. To achieve adaptive feature to different load time - series, cyclic iterative calculation isused to select the number of nearest vectors. According to unscented transformation theory, ensemble is formed by appropriate number of Sigma points of the predicted load value and data assimilation can be realized using ensemble Kalman filter. Therefore, optimum estimation can be obtained while measurement noise exists.Using non-instrusive measurement, forecasting results and erroranalysis for loads of electrical water heater and air -conditioner aswell as total household are done by adopting 50 end users datafrom Baoding of Hebei province as training data set. The resultsshow that, compared to those derived from pure phase-space reconstruction forecasting, the proposed method features have better-forecasting performance and good adaption both for individual appliance load and total household load.