Accurate load model significantly influences the numerical simulations of power system. The load model obtained by the component-based method does not adequately reflect the load time-varying characteristics, and the measurement-based method relies on fault data, but system faults rarely occur. Inspired by the theory of non-intrusive load monitoring, the load component decomposition and online modelling method based on the small disturbance response characteristic matching is proposed. An optimization model for the load component decomposition based on the curve fitting of small disturbance response characteristics is established. The construction method of a typical load model library is given and a neural network-based solution is designed. The proposed method can determine the load components that constitute the synthesis load model and their proportions by only analysing the small disturbance real-time measurement data. To verify the effectiveness of the proposed method, the distribution of the decomposition errors was analysed by the Monte Carlo simulation. Furthermore, the accuracy of the synthesis load model was verified by a fault simulation using the CEPRI 36-node system. Experimental results show that the proposed method can provide accurate load compositions and dynamically update the synthesis load model for the online power system simulation, planning and operation.