Load rebound characteristics (LRC), which means that restoration of distribution systems with large thermostatically controlled load (TCL) after an outage may create significantly higher load demand than normal (namely the demand if this outage does not occur), will make the reliability index—expected energy not supplied (EENS) of distribution systems obtained by traditional methods pessimistic and inaccurate. Since EENS is one of the base points of investment planning for distribution system operator, it is necessary to consider the LRC in the reliability evaluation and provide a more accurate evaluation on EENS. This paper proposed a new reliability evaluation method of smart distribution systems based on a nonnegative k-singular value decomposition (NN-K-SVD) algorithm and a sequential Monte Carlo (SMC) simulation. The NN-K-SVD algorithm is used to identify the TCL in total load. A reliability model considering the LRC is established based on an electric water heater model. The SMC simulation is used to mimic the operation of distribution systems based on the established reliability evaluation model, and calculate reliability indices. Furthermore, the feasibility of the proposed method is validated by extensive cases studies.